AI papers — 2026-08-10

Problem: Naively scaling up the number/diversity of multimodal training environments for agent learning does not reliably improve performance and can even degrade it, so the paper investigates how to design effective (not just large) environment distributions.

Method: The authors build a pool of 200 multimodal environments and empirically show mixed-environment training causes stronger negative transfer/gradient conflict in multimodal settings than in text-symbolic ones. They propose two techniques: (1) Ability-aware Environment Selection (AES), which decomposes agent trajectories (from Qwen3-VL-4B and Gemini-3-Flash) into atomic abilities via GPT-5 annotation, builds a meta-ability profile per environment, and greedily selects environments to maximize ability coverage while minimizing redundancy (profile similarity) and gradient conflict; and (2) Hierarchical Difficulty Curriculum (HDC), a two-axis curriculum with an outer axis of "harness weakening" (gradually removing textual scaffolds like text observations/states/hints/rules) and an inner axis of state-scale progression (e.g., grid size), each tracked per-environment.

Key results:

• Mixed-environment training drop: text-symbolic version -1.3% vs. multimodal version -10.7% (Table 1), confirming multimodal environments are more prone to negative transfer.
• Failure analysis of Qwen3-VL-4B (200 trajectories): 30% visual state extraction errors, 22-23% world modeling errors.
• Main results (Table 3, K=30 selected envs): AES achieves 95.6% average relative gain over base model vs. 43.4% for training on All Envs and 44.6% for Random-K.
• AES generalizes across model scale: on Qwen3-VL-8B, AES yields 144.3% relative gain on ID environments and 47.1% on OOD, despite being constructed using 4B trajectories/gradients.
• AES + HDC combined achieves 143.2% average relative gain over the base model across ID/OOD groups and both model sizes (best configuration in Table 3, e.g., 4B: ID-ST 45.0 vs. base 15.7/13.1, OOD-ST 22.1 vs. 13.8).
• Ablation on AES (Table 4): removing redundancy control drops OOD relative gain from 40.3% to 25.3%; removing conflict control drops it to 2.8%.
• Ablation on HDC (Table 5): scale-only curriculum gives 11.5% gain, harness-only gives 18.1%, combined HDC gives 27.7% (all relative to AES-only baseline).

Why it matters / caveats: The work shows environment "quantity" is not sufficient for multimodal agent RL and gives concrete, ability-grounded diversity/difficulty design principles. Authors note limitations: the environment pool is built from prior works rather than large-scale synthesis, experiments use a fixed compute budget so some environments may be undertrained, and AES's gradient-based conflict estimation adds nontrivial offline computation cost.

Problem: It is unclear why Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) behave so differently in multi-task LLM training — SFT typically requires mixed-data training to avoid forgetting, while RL is often trained multi-stage (one task per stage) without collapsing — and the underlying mechanism was previously unexplored.

Method: The authors empirically compare multi-stage and mixed-data SFT vs. GRPO-based RL on DeepSeek-R1-Distill-Qwen-1.5B/7B across four reasoning domains (math, science, logic, code), measuring parameter update magnitude (L2 norm) and pairwise cosine similarity between task-specific update vectors ΔW. They give a theoretical derivation showing SFT gradient interference is "norm-limited" (bounded by the absolute score-function magnitude, since SFT is off-policy toward a fixed expert distribution) while RL (GRPO) interference is "variance-limited" (bounded by intra-group rollout variance, due to the advantage function's zero-sum property and on-policy sampling). Building on the finding that RL updates across tasks are near-orthogonal, they propose Parallel-RL: train each task's RL independently and merge the resulting weight updates (via naive sum/mean, TIES, SVD rank-1, or an "Adapted" variant with light post-merge fine-tuning on 5% of data).

Key results:

• Multi-stage SFT causes a 23.1% average accuracy decline vs. base model; multi-stage RL achieves a 24.9% average gain (mixed-data SFT: +7.4%; mixed-data RL: +12.6%).
• Single-task training generalization: SFT improves its target task by +4.0% but degrades other tasks by -5.1% on average; RL improves target by +6.8% while also improving untrained tasks by +2.3%.
• Parameter update magnitude: average L2 norm of ΔW is ~3×10⁻² for RL vs. 7.4 for SFT (>2 orders of magnitude difference); RL updates are far sparser (~20% of parameters exceed magnitude 10⁻⁵ vs. 93% for SFT).
• Cross-task cosine similarity of ΔW: ~10⁻⁵ for RL vs. ~10⁻¹ to 1.0 for SFT (SFT sometimes has opposite-signed updates, e.g., Math vs. Code).
• On DeepSeek-R1-Distill-Qwen-1.5B (avg. across MATH500, AIME2025, MMLU, GPQA, KK, LiveCodeBench): Naive Parallel-RL(sum) retains 94.2% of single-task RL performance (+6.6% over base) vs. Naive Parallel-SFT(sum) retaining only 65.4% (-9.2% vs base); TIES Parallel-RL retains 97.4%; Adapted Parallel-RL achieves +10.7% over base, retaining 103.2% (surpassing single-task RL) with only 5% extra adaptation data. Similar pattern on the 7B model (Adapted Parallel-RL: +8.0%, 102.4% retention).
• Ablation: removing a task's ΔW from the Parallel-RL merge drops that task's accuracy by 7.1% on average while other tasks stay roughly stable (+0.6% on average), confirming decoupling.

Why it matters / caveats: The findings explain a widely observed empirical asymmetry between SFT and RL training recipes and motivate Parallel-RL as a more efficient, modular alternative to multi-stage or mixed-data RL. The authors note Parallel-RL is not merely model merging — it also requires task selection and per-task training tricks, and the smaller gains from mixed-data RL (vs. multi-stage RL) may stem from gradient imbalance across tasks, which is not the paper's focus.

Problem: Existing World-Action Models (WAMs) for end-to-end autonomous driving follow an "imagine-then-act" pipeline that generates future video frames at inference time to condition trajectory planning, which places costly video synthesis inside the real-time planning loop and substantially increases latency.

Method: SimWAM co-trains a pretrained video expert (a video Diffusion Transformer initialized from Wan2.2-5B, with its VAE and T5 text encoder) and a lightweight action expert (a DiT with hidden size 1024) jointly via flow matching, using an isolated attention mask so action tokens can attend to the current observation but never to future-frame tokens; this lets the entire video branch be discarded after training, leaving a self-contained planner that predicts trajectories directly from the current frame. The two experts share no parameters and interact only through shared attention, so the video backbone or action-expert size can be swapped/scaled independently. A reinforcement learning stage further optimizes the action expert by converting the deterministic flow ODE into a marginal-preserving SDE (following Flow-GRPO) and applying GRPO with the compositional NAVSIM PDM reward, updating only LoRA adapters, focused on the hardest navtrain scenarios.

Key results:

• SimWAM achieves 91.5 PDMS on the NAVSIM navtest benchmark using only a single front camera, a new state of the art, beating the strongest VLM-based planner SGDrive (91.1) by 0.4 points and world-model planners DriveLaW (89.1, +2.4) and DriveWAM (90.1, +1.4).
• Component analysis: action-only baseline reaches 86.6 PDMS; adding video co-training raises it to 90.3; adding RL raises it further to 91.5 (a 4.9-point total gain from co-training + RL).
• Isolated attention mask achieves the best PDMS (90.3) versus bidirectional (90.2) and action→video (90.1) attention variants, showing decoupling loses no benefit while enabling faster inference.
• Video backbone flexibility: Wan2.1-1.3B and Wan2.2-5B give comparable PDMS (90.2/90.3); a driving-pretrained Cosmos2.5 backbone reaches the best PDMS of 90.4, while a lightweight LTX-Video backbone drops to 88.7.
• Zero-shot transfer to nuScenes (no fine-tuning) achieves the lowest average collision rate of 0.04% among compared methods, with competitive average L2 error of 0.96m.
• Substantially lower inference latency than other WAM-based planners at comparable/better PDMS (shown in Fig. 1 latency-vs-PDMS comparison against DriveLaW, DriveVLA-W0, Epona, DriveWAM).

Why it matters / caveats: Demonstrates that video-generation supervision can be used purely as a training-time signal (not required at test time) to transfer traffic-dynamics priors into a compact planner, avoiding the latency cost that "imagine-then-act" WAMs pay in the real-time planning loop. The design is explicitly framed by the authors as "not claiming algorithmic superiority" but as a simple, solid, flexible baseline meant to benefit from future advances in video generation models.

Problem: Existing streaming-video benchmarks rely on short clips and multiple-choice questions, letting minimal baselines that use only the last few frames match complex streaming models, so they fail to jointly and rigorously evaluate real-time perception, long-horizon memory recall, proactive interaction, and tool use in continuous hour-scale video streams.

Method: The authors introduce StreamArena, a benchmark of 243 full-length videos (avg. 88.8 minutes) with 3,646 open-ended, human-annotated QA pairs (dual timestamped queries and evidence, validated by a three-stage annotation/cross-check/audit pipeline that removes ~27% of drafts) covering four capabilities: real-time perception (RTP), historical retrospection (HR), proactive interaction, and multimodal tool utilization. They then propose StreamMind, a two-tier agent architecture that separates a frontend (Front Worker for dispatch/answering, independently scheduled Monitor Workers for proactive alerts) from a backend (Memory Writer building a hierarchical event + entity-relation memory bank, plus Router/Recall/Search Workers for retrieval and external tool use), so latency-critical interaction is decoupled from asynchronous long-horizon memory construction and retrieval.

Key results:

• StreamMind ranks first among streaming systems on all four capabilities, improving over the strongest streaming baseline by 58.4% on RTP, 53.7% on HR, 228.1% on Tool use, and 54.7% on Proactive interaction (relative gains).
• Existing baselines show clear degradation patterns: AURA drops from 25.4% HR accuracy within 5 minutes to 10.5% beyond 30 minutes; VST (text-summary) gets 21.2% on retrospection; ThinkStream scores only 8.0/7.5/1.2/1.8% on perception/retrospection/proactive/tool-use respectively.
• Human reference performance: 91.8% RTP, 95.2% Tool, 91.5% Proactive, and HR accuracy of 80.7% with rewatching vs. 63.4% without (streaming condition), showing the task is hard even for humans under causal/streaming constraints.
• StreamMind improves every HR temporal-gap bucket (≤5min to >30min) over the best streaming baseline for that bucket, by 24.0% to 73.0% relative.
• Using the same Qwen3.5-397B-A17B backbone, StreamMind reduces pooled query-to-answer latency by 66.2% (81.4s → 27.5s) while retaining 89.7% of pooled accuracy versus offline query-triggered inference.
• Diagnostic ablation: ASR-only input reaches just 4.2% on RTP vs. 26.8% visual-only and 32.4% visual+ASR, confirming genuine multimodal grounding is required (not language shortcuts).

Why it matters / caveats: Provides an evaluation framework and reference architecture aimed at real deployment scenarios (embodied robots, wearables, always-on assistants) where prior short-clip/MC benchmarks were shown to be gameable. The authors note a remaining 63.4% accuracy drop from near-term to long-term retrospection (L2 to L4) showing long-horizon memory is not simply a capacity problem, and caution that comparisons across differently-scaled baseline backbones conflate model capacity with system design, so not all StreamMind gains should be attributed purely to the architecture.

Problem: Generic parameter-efficient fine-tuning (PEFT) methods developed for homogeneous Transformer stacks (e.g., LoRA) can silently fail or degrade accuracy when applied to real-time object detectors (YOLO family, RT-DETR, etc.), because these detectors have heterogeneous operators and detection-specific components (grouped/depthwise convolutions, DFL projections, deformable attention, MoE routing) that impose placement constraints absent in language models.

Method: The paper proposes YOLO-PEFT, which formulates adapter placement as an auditable constraint-planning problem over a detector's computation graph rather than proposing a new low-rank parameterization. A graph parser assigns each module an operator role and a semantic role plus interface/deployment metadata; a planner sequentially enforces operator-validity, detection-semantic, graph-interface, architecture-policy, and budget-feasibility constraints (each with a reason code), then does budget-aware rank assignment, and either emits a target-module plan or returns "Refuse" (falling back to Full-SFT) when reliability calibration predicts catastrophic ΔmAP. Accepted plans are lowered into a unified train-save-merge-export runtime supporting LoRA/RS-LoRA/DoRA/LoHa/LoKr/IA3/HRA/etc. and ONNX/TensorRT export, with merge-equivalence proofs for a fallback grouped-convolution LoRA backend.

Key results:

• Under the official VOC07+12 trainval→VOC07test protocol: planner-selected RS-LoRA reaches 0.7138 mAP50-95 on YOLO11s and 0.7307 on YOLO12s, vs. 0.6428 and 0.6662 for Full-SFT (gains of +7.1 and +6.5 points); best config (HRA) on YOLO12s reaches 0.7453 vs. 0.6662 Full-SFT (+7.9 points).
• On RT-DETR-L, all seven evaluated LoRA-family configurations fell below the catastrophic threshold (Δ<-0.05), a 7/7 (100%) unsafe-placement rate, triggering a calibrated Refuse→Full-SFT fallback.
• Controlled YOLO11 audit: LoRA reduces peak training VRAM by 43.9% (28.57→16.03 GB) but takes 1.72x longer training time (118.2s→203.8s).
• Catastrophe rate (ΔmAP<-0.05) rises with attention ratio φ_attn: 0/10 on CNN-only YOLO11n, 6/7 on YOLO12n, 7/7 on RT-DETR-L.
• Ablation on YOLO12s: naive placement without constraints yields 0.6900 mAP; adding operator-validity filtering raises it to 0.7094; adding detection-head semantic exclusions raises it to 0.7307 (largest single stability gain).
• Rank sensitivity: r=8→0.7288, r=16→0.7307, r=32→0.7363 (diminishing returns), supporting r=16 as a Pareto default.
• MoE stress test (YOLO-Master-EsMoE-S, planner/expert-targeting disabled): HRA reaches 0.7454 mAP50-95 vs. logged Full-SFT 0.6891; BOFT and OFT variants failed to complete/produce metrics.
• DoRA without RS-LoRA scaling collapses to 0.6112 mAP (vs. 0.7307 with RS-LoRA), showing training-prior interactions are non-additive.

Why it matters / caveats: YOLO-PEFT replaces manual, trial-and-error target-module selection with explicit, auditable pre-training risk planning, and can save costly failed training runs (e.g., ~23.8 minutes across seven avoided RT-DETR-L attempts in their protocol-local estimate). The authors explicitly caveat that refusal calibration is validated only within observed architecture families (leave-one-variant-out, not leave-one-family-out), so generalization/safety on unseen detector architectures, COCO/domain-shift settings, and expert-aware MoE planning remain open problems.

Problem: Existing RLVR reward designs for audio reasoning are complementary but limited: outcome-based rewards supervise only the final answer (letting models ignore the audio), while process-based rewards use coarse, hand-crafted, fixed criteria that don't adapt per-question or stay grounded in acoustic evidence, and any static rubric saturates as the policy improves.

Method: The paper introduces AUDIORUBRICS, an RL framework built on GRPO that uses a single audio-capable model (Gemini-3.1-Pro) as both rubric generator and judge to synthesize per-question, audio-grounded rubrics directly from the raw waveform (not a transcript). At each training step, rubrics are elicited and judged against the group's sampled rollouts, non-discriminative rubrics (identical verdicts across all rollouts, i.e., zero within-group standard deviation) are pruned, new rubrics are elicited by contrasting stronger vs. weaker rollouts, and survivors are re-weighted — producing an evolving rubric reward Rrub combined with an outcome reward (accuracy + format) and an overthinking penalty that linearly penalizes reasoning-trace length. Total reward: R = Rout + γ·Rrub + δ·Rover, optimized via GRPO on Qwen2.5-Omni-7B using the AVQA-derived dataset (40,176 samples).

Key results:

• Overall accuracy: AUDIORUBRICS achieves best-in-class results among similar-size models on MMAU Test-mini (78.00%), MMAR (65.80%), and MMSU (65.86% overall; 52.75% avg across perception/reasoning).
• On MMSU perception, AUDIORUBRICS beats the best similar-size baseline (CESAR, 60.16) by 4.3 points (66.14 vs. 60.16), an 8.9% relative improvement.
• Ablation (Table 6): full method 78.00/65.80/65.86 (MMAU/MMAR/MMSU) vs. RL-training-only baseline (no rubrics) 75.20/62.20/63.14, and vs. no-RL base model 65.20/56.70/60.57 — each component (static rubrics, evolving rubrics, length penalty) adds incremental gains.
• Reward-weight sweep: rubric weight γ=0.5 is optimal, giving +3.72%/+5.79%/+4.31% relative gains over plain GRPO on MMAU/MMAR/MMSU; overthinking penalty δ=0.15 is optimal (+3.72%/+5.79%/+4.31%), with too-small δ causing overlong traces and too-large δ over-compressing reasoning.
• Judge/generator ablation: replacing Gemini-3.1-Pro with the weaker GPT-audio-1.5 as rubric generator/judge causes performance to drop even below the vanilla GRPO baseline (e.g., 65.80/61.30 vs. 78.00/65.80 on MMAU/MMSU), showing gains scale with generator/judge capability.
• Newly-adopted-rubric ratio grows from ~35% early in training to ~60% by the end, confirming the rubric set continually updates rather than staying static.

Why it matters / caveats: By supervising the reasoning process with per-question, audio-grounded, and self-evolving criteria rather than a fixed rubric or outcome-only signal, the method avoids reward saturation and measurably improves genuine audio-perception grounding (not just final-answer accuracy). The authors note a caveat: the approach depends on a sufficiently capable rubric generator/judge model — a weaker judge (GPT-audio-1.5) produces noisy, poorly grounded rubrics that can perform worse than plain GRPO — and the rubric reward requires careful weighting (γ) alongside a length penalty (δ) to avoid reward hacking via overlong, redundant reasoning traces.

Problem: Existing world-model-based navigation methods score candidate trajectories using only one or a few point predictions of future states, which is inadequate for large-scale outdoor UAV navigation where substantial future-state uncertainty (e.g., from occlusion or long-horizon drift) makes such scoring unreliable.

Method: The paper proposes UA-NWM (Uncertainty-Aware Navigation World Model), which formulates trajectory scoring as conditional out-of-distribution detection: it predicts future DINOv3 latent features with a lightweight deterministic backbone (context length 4, horizon 8, 32 latent tokens, built on a DeltaWorld-style delta-token transition model) and introduces a Hierarchical Error Projection (HEP) module that decomposes the prediction-goal discrepancy into a parallel, uncertainty-explainable component and an orthogonal, unexplained residual, using only the residual for trajectory cost. HEP builds a coarse-to-fine pyramid over grid scales {1,2,7,14}, fitting a rank-R (R=2 best) low-rank subspace per cell via ridge-regularized least squares to model plausible future variation. The authors also construct a new benchmark, AirGoal-10k (9000 train / 1000 val / 1000 test trajectories in AirSim), for large-scale outdoor 3D UAV image-goal navigation.

Key results:

• Trajectory ranking on AirGoal-10k (32 candidates): UA-NWM achieves ATE 1.09 / RPE 0.30, vs. NWM 1.37/0.37, RAE-NWM 1.50/0.41, One-Step WM 1.44/0.39, at 8.47 ms/frame (vs. NWM's 438.67 ms and RAE-NWM's 386.00 ms).
• Standalone CEM planning: UA-NWM ATE 1.22 / RPE 0.33, best among all policy- and world-model-based baselines (e.g., NWM 1.40/0.39, GNM 1.29/0.35).
• Online closed-loop navigation (100 AirSim episodes): UA-NWM reaches 76.0% Success Rate / 64.5% SPL, vs. RAE-NWM 70.0%/57.8%, MWM 69.0%/56.0%, NWM 63.0%/52.1%, while planning in 2.70 s/step vs. NWM's 191.71 s/step and RAE-NWM's 202.54 s/step.
• Ablations: removing any HEP pyramid scale degrades performance (finest scale-14 contributes most); rank R=2 is optimal (R=1 underfits, larger R absorbs true errors).
• Real-world quadrotor deployment (MacBook Air + Jetson Orin NX) succeeded on five navigation tasks across multiple scenes with zero-shot sim-to-real transfer, ~9 seconds per on-device planning step.

Why it matters / caveats: By representing an uncertainty subspace instead of sampling multiple futures, UA-NWM achieves both stochastic-model-level robustness and near-deterministic-model inference speed, and real-world flight tests support practical deployability without real-data fine-tuning. The authors note that modeling uncertainty too broadly (as in NWM/RAE-NWM) can absorb genuine trajectory errors and weaken scoring discrimination, which is why the hierarchical, region-specific formulation is needed.

Problem: Activation Oracles (AOs) — language models trained to answer natural-language questions about another model's internal activations — are themselves learned systems, and it is unknown whether fine-tuning an AO on a subject model that hides a concept makes the AO a better "specialist reader" of that concept or instead causes it to fail to report it.

Method: The authors use a controlled Taboo Word Guessing setup (following Cywinski et al. 2025) where Qwen3-8B subject models are LoRA fine-tuned to internally use one of five hidden words (leaf, moon, wave, flag, book) under "cooperative" (give hints, don't say the word) or "strict" (refuse to reveal) protocols, at two Taboo-data mixture strengths (α=0.5, 1.0). They then train Activation Oracles (LoRA-fine-tuned copies of the backbone, following Karvonen et al. 2025) on residual-stream activations (layer 18) from these subjects, comparing a base AO (trained on the unmodified subject), "own" FT-AOs (trained on a subject sharing the test subject's hidden concept), and "cross" FT-AOs (trained on a subject with a different hidden concept). They use linear probes, a ∆-LogitLens diagnostic, and layer-range LoRA ablations to localize where recovery fails.

Key results:

• Base AO recovery of cooperative subjects in the HINT regime: 93% exact recovery, 95% semantic recovery, with the target essentially top-ranked.
• For cooperative subjects at α=1.0: exact recovery drops from 94% (base AO) and 71% (cross FT-AO) to 35% (own FT-AO); semantic recovery drops from 99% and 86% to 51%.
• Linear probes on subject activations achieve ≥98% mean accuracy (vs. 16.7% chance) for identifying the hidden concept across all regimes/layers tested, including SAMETEXT and OFFTOPIC (no relevant behavioral content).
• Concept-specific blind spots persist even for multi-concept subjects (trained on both "leaf" and "moon"): leaf-AOs assign low probability to "leaf" but higher probability to "moon," and vice versa for moon-AOs.
• Layer-range ablation: removing FT-AO LoRA updates in layers 18–23 gives the strongest and most consistent restoration of target probability; ablating early or final layers does not.

Why it matters / caveats: The findings show behavioral leakage, representation-level decodability, and AO-verbalizability can come apart — a learned activation-reading interface can acquire training-induced blind spots rather than becoming a more faithful reader, raising a reliability concern for using AOs as interpretability/auditing tools. The authors note limitations: results come from a single backbone (Qwen3-8B), a small set of synthetic Taboo concepts (not naturally occurring hidden knowledge like backdoors), and LoRA-trained subjects/oracles, so generalization across model families, scales, and more realistic hidden-information settings remains untested.

Problem: Knowledge distillation is the standard (and expensive) way to recover quality in a compact LLM derived from a larger teacher, but online distillation requires keeping the teacher resident in memory and recomputing its forward pass every step, and the resulting dense vocabulary-sized logit tensor for the KL loss creates a memory spike that caps the trainable context length.

Method: The paper presents two systems contributions plus supporting ablations, evaluated with an ~8B Llama 3.1 Instruct teacher and a compact ~3.2B student (Minitron-style pruned initialization). First, offline distillation: precompute and cache only the teacher's top-K=100 logits per token once, then train the student against this cache instead of running the teacher live. Second, a fused, chunked KL loss that extends memory-efficient cross-entropy kernels (like Cut Cross-Entropy/Liger) to the knowledge-distillation setting: it processes hidden states chunk-by-chunk, fuses the output projection into the loss so the full [S,B,V] logit tensor is never materialized, and uses a closed-form gradient (M·softmax minus sparse teacher correction) recomputed on the backward pass, making peak memory linear rather than O(sequence_length × vocab_size).

Key results:

• Offline vs. online distillation reach near-identical training-loss curves despite offline training against only top-100 cached logits; offline lowers peak memory from ~103GB to 78GB, runs ~29% faster per iteration (25.9s→18.5s), and raises throughput from 237 to 331 TFLOP/s (~40% higher; abstract states up to 41%).
• Among offline KL implementations at 8K context on one H200: full dense KL peaks at 78GB, forward-chunked at 62GB, fused chunked at 58GB, all producing the same training-loss curve.
• The fused chunked loss enables training at 32,768 tokens (~4x the context of dense KL) on a single H200, where dense KL would need ~250GB (exceeding the 141GB H200 capacity) versus ~128GB for the fused loss.
• Scaling to GPT-OSS-20B at 32K context on 8×H200 nodes: fused loss lets the model fit on 1 node instead of 4, cutting step time from 57.0s to 12.23s (~5x faster) and raising per-GPU throughput from 74.2 to 345.7 TFLOP/s.
• Isolated toy-network loss-kernel benchmark: at 32K tokens, peak memory is 85.2GiB (dense), 17.7GiB (forward-chunked), 5.45GiB (fully chunked) — a 15.6x reduction; dense KL fails (OOM) from 64K tokens onward; at 256K tokens the fully chunked loss uses 11.6GiB vs 134.2GiB for forward-chunked, and is 3.3x faster in iterations/second (0.630 vs 0.190) despite being slower than forward-chunked at 32K.
• Ablations: logit-level KL is indispensable (feature loss alone collapses the student to ~28% MMLU/~4% GSM8K); logit KL alone recovers 59.9% MMLU/65.9% GSM8K, and adding a hidden-state feature loss improves this to 60.6%/67.5%. Naive sequence packing (all-ones attention mask) costs only about one point of MMLU versus non-packed training.

Why it matters / caveats: Makes large-scale distillation "healing" and hundreds of ablation runs practically affordable by removing the teacher from the training loop and removing the logit-tensor memory spike that otherwise caps context length — directly enabling long-context distillation on commodity single-GPU setups. The authors note the study is limited to a single teacher-student pair (8B→3.2B) and one hardware/framework stack (Megatron-Bridge/ModelOpt on H200s), so generalization to other model families, compression methods, student sizes, or hardware is untested; the toy-network loss-kernel benchmark also does not capture end-to-end training speed, convergence, or interactions with attention/optimizer state at long lengths.

Problem: Existing MLLM-based multimodal embedding models cannot jointly satisfy the industrial requirements of billion-scale indexing efficiency and fine-grained semantic discrimination — contrastive models are fast but rely on coarse pair-level supervision, while chain-of-thought (CoT)-based reasoning embedders improve discrimination but require expensive explicit generation impractical for online serving.

Method: The authors present Douyin Multimodal Embedding (DME), a two-stage bi-encoder trained on a generative MLLM backbone (built on Qwen3-VL/Qwen3.5-family backbones at 2B and 9B scale). Stage 1 performs large-scale contrastive pre-training (~25M query-document pairs) using a simple <emb> readout token to build a broad multimodal embedding space across text, image, video, and visual-document modalities. Stage 2 adds two mechanisms trained on ~5M higher-quality examples with teacher-generated (Seed-2.0-Pro) structured supervision: Evidence-Grounded Typed Latent Reasoning, which uses modality-aware anchor tokens and typed hidden-space latent states (localize, align_pos, reject_neg, summarize) to ground the embedding in retrieval-relevant evidence within a single forward pass; and Cross-Conditional Reconstruction, which uses the query/document embedding as a decoding prefix for Next-Token-Prediction and Multi-Token-Prediction reconstruction of the counterpart's text — used only during training so inference remains a standard single-pass dense encoder.

Key results:

• On MMEB-v2, DME-2B achieves an overall score of 74.8 and DME-9B achieves 78.4, described as state-of-the-art among comparable-scale models, with particularly strong gains on video and visual-document retrieval.
• In Douyin's production in-house offline evaluation set, DME delivers a 2.92% relative improvement in overall score, with consistent gains across all cross-modal retrieval directions.
• Online A/B testing on Douyin search shows a 0.1% Lifetime (LT) gain.
• Latent reasoning tokens are reported to add only marginal query-side encoding latency overhead versus a standard bi-encoder.

Why it matters / caveats: DME shows that reasoning-style, evidence-grounded and generatively-supervised training can be baked into a bi-encoder at training time only, giving fine-grained discrimination without sacrificing the low-latency single-pass inference needed for billion-scale industrial retrieval (deployed across Douyin generative search, image search, and AI search); the reported online production gain (0.1% LT) is comparatively modest relative to the large offline benchmark gains, suggesting real-world impact is smaller than benchmark numbers imply.

Problem: Standard additive (Chinchilla-style) neural scaling laws assume model size (N) and training data (D) affect loss independently (forcing the cross-derivative ∂²L/∂N∂D to be exactly zero), which causes systematic under- and overestimation of loss at data-scarce and overtraining extremes of the training grid.

Method: The authors introduce the Skaling law, L(N,D) = (A/N^α + B/D^β)^k + E, which keeps Chinchilla's additive inner terms and independent exponents (α, β) but raises their sum to a single free outer coupling exponent k (recovering Chinchilla exactly at k=1), reinstating a Kaplan-style N–D interaction with just one extra parameter. They first empirically validate the coupling by estimating loss derivatives via moving-least-squares (MLS) and Gaussian-process (GP) surrogates on training grids (Farseer: 404 configurations, 100M–6.4B params, 1B–512B tokens; and their own SK-Grid: 134 configurations, 134M–4.9B params, 316M–316B tokens), confirming a non-zero, predominantly negative mixed derivative. They also propose an "L-shape" sparse sampling strategy that trains only along the cheap N-band and D-band edges of the grid rather than the full square.

Key results:

• Skaling reduces MAPE by 1.5–3× versus Chinchilla; on the full Farseer grid, single-axis extrapolation MAPE falls from 1.48% to 0.47% (Ext. N) and 1.98% to 0.88% (Ext. D); on SK-Grid, far-extrapolation MAPE drops from 5.17% to 0.70% (full grid) and from 14.63% to 1.15% (L-shape grid).
• Skaling wins at 76% of (N,D) configurations in the boundary-error comparison (median 2.2× more accurate, ≥4× at a third of configurations).
• Using the sparse L-shape grid (~10× less compute), Skaling remains near or better than full-grid Chinchilla, while Chinchilla itself degrades substantially under the same restriction (Farseer interpolation MAPE rises from 0.77% to 2.51%; SK-Grid far-extrapolation MAPE rises from 5.17% to 14.63%).
• Fitted coupling exponent k is consistently sub-unit (k≈0.31–0.45 across grids), and the fitted irreducible loss E drops sharply (e.g., 0.45→0.03 on Farseer full grid), indicating a genuinely coupled, non-additive surface.
• In compute-extrapolation experiments predicting the 112 highest-compute Farseer runs from cheap ones, Skaling achieves pooled MAPE 0.60±0.27%, a 3.9× reduction over Chinchilla (2.34%) and better than the more heavily parameterized 9-parameter Farseer law.
• The Chinchilla-predicted optimal token-to-parameter ratio diverges from Skaling's by up to ~100× at frontier compute scales; empirical (non-parametric) gradient estimates of the optimal ratio agree with Skaling's exponent (-0.11 to -0.15) and disagree in sign with Chinchilla's near-flat prediction (+0.03).

Why it matters / caveats: A single added exponent meaningfully improves prediction accuracy exactly where it matters most for compute-budget planning — at data-scarce/overtraining extremes and for extrapolating to frontier scales — and enables much cheaper profiling via sparse L-shape grids; the authors note the benefit shrinks when the underlying data is closer to additive (e.g., k≈0.77–0.90 on Farseer-code and original Chinchilla data, where Skaling performs at roughly Chinchilla-level accuracy), and that the direction of the coupling's effect on optimal token-to-parameter allocation is dataset/architecture-specific rather than universal.

Problem: In autoregressive video world models, once generation rolls out beyond the training context length, past visual content stored in the KV cache becomes effectively unreachable (not addressable) because temporal RoPE offsets fall outside the training distribution, and naive cache compression further corrupts that content via RoPE phase cancellation.

Method: The paper proposes WorldTrace, a training-free memory framework that assigns each compressed "summary slot" a distinct, in-distribution virtual position based on slot rank (not absolute timestamp or a capped/centroid position), keeping every slot addressable at any generation horizon. Two content writers instantiate this: WorldTrace-Field, which averages keys in canonical (unrotated) space before re-rotating to the slot's virtual position (for temporal coherence), and WorldTrace-Landmark, which freezes canonical keys of detected scene-entry frames verbatim at their virtual position (for episodic recall). The authors also frame memory compression as structured sparse attention/nonnegative matrix factorization to justify the two writer designs, and introduce LoopBench, a benchmark of controlled scene-revisit loop trajectories (ABA, ABCA, ABCDA topologies).

Key results:

• WorldTrace-Field improves temporal consistency (TempSSIM) by +15.5% relative at N=48 chunks versus sliding-window baseline, and is best on both TempSSIM and Scene Drift at N=48.
• At N=8/N=16, WorldTrace-Field's position assignment outperforms Block-relative positioning by +5.9%/+2.8% TempSSIM and Centroid-linear by +9.5%/+13.8%.
• WorldTrace-Landmark improves episodic recall (PAC metric) over sliding-window baseline across all tested topologies/horizons/orientations/multi-revisits, e.g., ABA (N=16): 0.864 vs 0.723; ABA long (N=32): 0.825 vs 0.627.
• At extended horizon N=256, verbatim landmark recall sustains PAC≈0.99 versus 0.610 for canonical-K anchoring.
• Canonical (vs. naive RoPE-space) key averaging reduces LatentDiff from 0.312 to 0.233 at Ns=4 (−25.3%).
• Overall reported headline gains: WorldTrace-Field +15.5% temporal consistency; WorldTrace-Landmark +19.5% episodic recall on LoopBench ABA loops.

Why it matters / caveats: This is a training-free, drop-in fix for a real failure mode (loss of scene persistence) in interactive video world models used for game engines and robot simulators. The authors note WorldTrace is designed specifically for temporal-RoPE autoregressive models with a fixed KV-cache budget; WorldTrace-Field can blur fine scene details while WorldTrace-Landmark depends on correctly detecting and retaining the right scene-entry events, and both are described as simple structured approximations rather than adaptive/learned solutions.

Problem: Existing test-time-training (TTT) sequence-model variants are each hard-coded as one-off implementations, which makes it difficult to systematically design new TTT variants or isolate which individual design choices (fast-weight network, loss, learning rate, decay, normalization) are actually responsible for observed performance.

Method: The paper introduces Modular TTT, a framework that represents the TTT inner learner as a directed acyclic graph (DAG) of primitive operators (Linear, Gate, Norm, Act, Add, Mul), each registered with explicit train-view forward, train-view backward, and causal query-view rules; given any learner DAG, these primitive rules are automatically composed into the full graph-level TTT computation (including the fast-weight state transition), removing the need to hand-derive a new global update rule per variant. Using this framework they run controlled ablations (at 160M/410M scale, 10B tokens) over loss function (MSE, inner-product, L1, RMSE), learning-rate initialization, weight decay (none/scalar/vector), nonlinearity (GELU/SiLU/Norm), and fast-weight network depth, implemented in PyTorch within the Flame framework and evaluated with lm-eval-harness. They then scale the selected best variant to 410M and 1.45B parameters trained on 100B tokens, benchmarking against LLaMA, Gated DeltaNet (GDN), and LaCT.

Key results:

• Loss ablation: MSE (410M loss 2.7949) and inner-product (2.7938) are competitive; L1 (2.9727) and RMSE (2.8041) perform substantially worse.
• Learning-rate init: "small-lr init" (η0≈10⁻³) beats standard init (η0≈1) consistently, e.g., MSE+scalar-decay 410M loss drops from 3.0820 (standard) to 2.7949 (small-lr init).
• Decay: scalar decay recovers most of the gain of vector decay (410M loss 2.7949 vs. 2.7821) at far lower cost — vector decay costs ~25% lower throughput and ~3GB more peak memory.
• Nonlinearity: Linear+SiLU (410M loss 2.7849, MSE) and Linear+GELU (2.7844) beat plain Linear (2.7949); Norm is mixed/less stable (2.8192).
• Deep memory (2+ layer fast-weight MLPs) never surpasses the shallow single-linear-layer frontier (best deep variant loss 3.1144 vs. shallow Linear-SiLU 3.0205 at 160M).
• Efficiency: analytic backward operators give 1.65x (Linear) and 2.62x (Norm) speedup over autograd reference; end-to-end, Modular TTT achieves ~2.2x–3.3x training throughput over the official TTT implementation (e.g., TTT-MLP: 21,336.9 → 71,101.1 tokens/GPU/s).
• At scale (410M/1.45B, 100B tokens), the selected Modular TTT variants achieve training loss and downstream multiple-choice accuracy comparable to GDN (e.g., 1.45B: Modular TTT M-SiLU multiple-choice avg 55.67% vs. GDN 56.30%, LLaMA 55.67%), though containment-style tasks (SWDE/SQuAD/FDA) and long-context RULER recall lag notably behind LLaMA, especially at 8k context.

Why it matters / caveats: The framework turns TTT design into a systematically explorable modular space rather than a set of custom implementations, both accelerating experimentation and clarifying which components (small-lr init, decay, single-layer nonlinearity) actually matter versus which don't (deeper networks, normalization, residual/gating connections give little benefit); the authors themselves note precise long-context recall remains a limitation of the shallow Modular TTT variants relative to LLaMA.

Problem: It is unclear whether AI-generated code exhibits a distinct, measurable quality/performance profile once it survives review and reaches production at large industrial scale, since most prior studies evaluate generated code only in controlled prompting/benchmark settings rather than in the full production lifecycle.

Method: The authors conduct a large-scale empirical study inside a large tech company (Google), using authoring-time provenance to track AI-generated vs. human-written C++ across 3.52 million submitted changes (April 2025-April 2026), joining a taxonomy of static-analysis findings (5 quality attributes: Efficiency/Resource Use, Correctness/Safety, Maintainability/Readability, Modernity/API Evolution, Policy/Portability/Environment Fit) with code-review, reliability, and post-deployment compute/memory monitoring data. They also run an intervention study (RQ4) testing "taxonomy-informed feedback" — reimplementing 50 benchmark C++ functions under 3 prompt stages (base, general feedback, category-specific feedback) with 3 runs each (450 generations total).

Key results:

• AI-generated code's share of submitted code (all languages) rose from 28.99% (Apr 2025) to 68.62% (Mar 2026); majority-AI C++ changes grew from 27.65% to 59.69% over the same window.
• Static issue-rate gap concentrated in two categories: Interface and Coupling Burden (rate ratio AI/human 1.15) and Copy and Allocation Overhead (1.39), which together represent 82.21% of the total positive absolute rate gap; Efficiency/Resource Use attribute ratio 1.23, Maintainability/Readability 1.08.
• Source-level efficiency: AI code uses ~2.0x more explicit loops, ~0.4x (30-40% less) standard-library/API calls, ~2.0x more map-access and container-insertion (use-emplace) warnings, and 1.39x more missing-move warnings.
• Compute/memory: AI-heavy functions show ~5% higher relative compute growth and ~8% higher relative memory growth vs. human-written functions by early 2026 (1.31x vs 1.25x compute baseline; 1.36x vs 1.25x memory baseline).
• Review friction: AI-generated changes receive 1.92x as many blocking review comments and 1.39x as many total comments; reviewer iterations 1.24x, time-to-merge 1.19x, submit attempts 1.07x; build-failure rate ~1.3x; but revert rate is actually lower (~0.9x) than human-written code.
• Intervention (RQ4): category-specific feedback (Stage 3) reduced targeted static findings by 11.1% relative to Stage 1 (and 12.5% relative to baseline) and improved the efficiency score Reff from 0.294 (Stage 1) to 0.385 (Stage 3), a 31% improvement.

Why it matters / caveats: The study argues pass-rate/completion metrics are insufficient — production quality costs show up as chronic maintenance/compute burden rather than acute failures/reverts, and this burden is concentrated enough (2 categories driving most of the gap) to be tractable for tooling intervention. The authors note the findings are C++- and organization-specific (one large monorepo), the analysis is observational/correlational (not causal), individual model versions could not be disentangled due to anonymization, and reviewers showed no clear ability to catch these localized inefficiencies via standard review metrics, motivating automated upstream interventions.

Problem: Autoregressive/diffusion-based dynamical-system surrogates accumulate rollout error over long horizons, but at deployment there is no ground truth available to measure that error, and standard uncertainty methods (ensembles, dropout, sample dispersion) measure model self-disagreement rather than actual error, degrading under the very distribution shift that rollouts induce.

Method: The authors train a single conditional latent diffusion model (a diffusion transformer denoiser, conditioned via adaLN-Zero) that steps a dynamical system forward or backward in time via a direction flag c_d ∈ {+1,-1}, using the same shared weights for both directions. At test time they compute round-trip consistency error C_i by rolling i steps forward then i steps backward and measuring the discrepancy from the original starting state — a self-supervised, ground-truth-free proxy for the true (unobservable) rollout error E_i. They provide a Lipschitz/co-Lipschitz theoretical certificate (Proposition 1, Corollary 1) bounding when small C_i implies small true error, and fit a heteroscedastic Gaussian calibrator (on log C_i) to convert the raw signal into a calibrated error-magnitude prediction.

Key results:

• On held-out MHD trajectories: C_i ranks rollout error with Spearman correlation 0.91–0.98 across trajectories at fixed depth, and 0.69±0.16 within trajectories.
• The calibrator predicts error magnitude within a factor of 1.14 (68% coverage) and 1.29 (95% coverage), a 1.0-nat improvement in held-out log-likelihood over a depth-only baseline; gains hold at every probed depth (0.6–1.3 nats).
• Flags the out-of-distribution Orszag-Tang vortex with AUROC 1.0 at depths 5–10 (0.98 trajectory-mean), exactly where the sampling-dispersion baseline inverts (AUROC 0.00 at depth 5).
• Selective prediction: deferring the riskiest 20% of predictions cuts incurred error by 15% (vs. only 5% for a depth-only deferral baseline) — roughly 3x better.
• Bidirectional training beats direction-specialist models in both directions by 7–10% despite seeing half the per-direction data (e.g., 0.0630±0.0006 vs 0.0675±0.0007 forward).
• On LE-PDE-UQ's Navier-Stokes benchmark, the single bidirectional model reaches block-relative L2 of 0.244, within 1.16x of their best single-model-with-uncertainty (0.211) and 1.29x of their 10-model ensemble (0.190), at roughly 1/10th the training cost; composing μ(C) with seed-spread shape gives the best training-free calibration (MA=0.082) vs. their learned single-model σ (0.058) and 10-model ensemble (0.014).
• On CelebV-HQ face video, C ranks clips by error with Spearman 0.79–0.81 (latent) and predicts held-out error within 1.28x (68% coverage, 95.3% actual coverage).

Why it matters / caveats: Gives a "free" (one extra rollout, no ensembles/retraining/ground truth) trust signal usable as an early-stop/deferral criterion for deployed scientific surrogates (plasma control, accelerator diagnostics, weather). The authors explicitly note the check is necessary but not sufficient (forward/backward errors could cancel, though empirically rare — 95.1% of points fall within calibrated bounds), the signal degrades at very deep rollouts and for ill-posed/strongly dissipative backward dynamics, and it lives in latent space rather than physically calibrated field-space.

Problem: Existing LLM-based optimization systems for prompts, programs, and ML workflows rely on external hand-coded controllers (evolutionary search, bandits, textual-gradient methods) that dictate the search policy while the LLM only proposes local edits, raising the question of how much of that search policy could instead be internalized by the LLM agent itself.

Method: The authors present ReASearch (Reasoning-driven Agentic Search), a controller-light framework where a single tool-using code agent (with Python/Bash execution, file I/O, and a persistent "lessons.md" memory file) autonomously decides what to evaluate, how to diagnose failures, which edits to make, and when to verify or restart, using the same task-agnostic agent loop across three domains. Domain-specific tools expose the core optimization operations (e.g., get_next_minibatch, call_student_model_batch, validate_candidate for prompts; evaluate and edit_code for programs; run_experiment for ML workflows) but no external algorithm governs candidate selection or budget allocation — that logic emerges from the agent's own reasoning. The backbone models tested include Claude Sonnet 4.6, GPT-5, and open-source models (GLM-5, Kimi-2.5).

Key results:

• Prompt optimization vs. GEPA (test-set accuracy): AIME 52.00% vs 50.67% (baseline 46.00%); GSM8K 83.40% vs 82.11%; HotpotQA 67.60% vs 65.80%; Terminal-Bench 2.0 53.33% vs 42.22% (baseline 35.56%).
• Program evolution: on Circle Packing and Heilbronn Triangle tasks ReASearch (Sonnet 4.6) matches or exceeds prior human best-known results (e.g., Heilbronn n=12: 0.03260 matching best-known, vs AdaEvolve's 0.02552); on ARC-AGI-2, ReASearch/Sonnet 4.6 reaches 85.0% train / 50.0% test accuracy vs AdaEvolve's 21.9%/12.5%; on systems programming, EPLB score improved from a 0.21 plateau to 0.23, and a transaction-scheduling agent cut code from 1752 to 244 lines for a +51% improvement.
• ML workflow optimization vs. Claude Code baseline: IMG-100 83.99% vs 78.59% accuracy; Atari Q*bert reward 4500 vs 1250; MuJoCo reward 5267 vs 3986; Crypto correlation 0.1110 (Kaggle rank 6) vs 0.0999 (rank 29), improving from a rank-36 baseline to rank 6 within 15 experiments; NanoGPT results were statistically indistinguishable (0.976 vs 0.974 BPB).
• Overall gains reported as 2% to 40% over strong domain-specific baselines across 14 tasks; ablations show removing memory or the Python execution tool degrades performance (e.g., AIME drops from 52.00 to 49.33 without memory; ARC-AGI-2 test accuracy drops from 50.0% to 32.5% without Python tools).

Why it matters / caveats: The results suggest that with sufficiently capable, rich-feedback tool-using agents, complex optimizer behaviors (verification, reverting, hypothesis-driven revision, budget allocation) can emerge purely from reasoning rather than requiring hand-engineered search algorithms, potentially simplifying and generalizing optimization pipelines; the authors note performance scales with the backbone model's agentic capability (open-source backbones like GLM-5/Kimi-2.5 underperform Claude Sonnet 4.6 somewhat, especially on harder tasks like Terminal-Bench 2.0) and that costs remain non-trivial (a full prompt-optimization run costs under $20 in API usage).

Problem: In privileged on-policy distillation for multi-turn agents, a synchronized teacher re-scores the student using training-only references (e.g., successful trajectories), but as the student's own actions diverge from the reference and it reaches states the reference doesn't cover, applying that reference indiscriminately creates "state–reference mismatch" that can actively degrade guidance rather than help it.

Method: The authors propose State-Matched Routing and Contextualized Self-Distillation (SMRC-SD), which at each turn reconstructs a compact state signature for the student's current execution state and matches it against pre-action state signatures reconstructed from a same-task successful reference trajectory; self-distillation loss (SDL) is applied only on turns where a match exists (state-matched routing), and for matched turns the teacher context is built from the full reference path plus a current-state summary plus the specific grounded continuation action (contextualized self-distillation), rather than feeding the teacher the full raw reference path unconditionally as the "FullPath-SD" baseline does. The method is built on top of GRPO with a chosen-token K3 distillation estimator (from Skill-SD/SDAR), and is evaluated on Qwen2.5-3B-Instruct and Qwen3-1.7B across ALFWorld and WebShop.

Key results:

• With Qwen3-1.7B, SMRC-SD improves ALFWorld task success (Average@4) from 0.746 (FullPath-SD) to 0.865, and Pass@4 from 0.836 to 0.914; WebShop Score improves from 0.694 to 0.825 and Acc from 0.574 to 0.693.
• With Qwen2.5-3B-Instruct, SMRC-SD improves ALFWorld Average@4 from 0.766 to 0.883 (Pass@4: 0.852→0.938) and exceeds the reported SDAR ALFWorld Average@4 of 0.844; WebShop Score improves from 0.842 to 0.863, Acc from 0.734 to 0.736.
• Ablation (Table 3): adding matched-only routing alone (fixed plain-path context) raises ALFWorld Avg@4 from 0.746 to 0.836; adding dynamic/state-contextualized guidance on top raises it further to 0.865; dynamic context without routing only reaches 0.695 — routing is the dominant contributor, contextualization adds an extra gain.
• Turn-identity control: matched-turn selection (0.836) beats a random-turn selection of the same count (0.723), showing the gain isn't from mere sparsity.
• Structured-state matcher vs. history matcher: expands candidate coverage from 15.4% to 20.2%, recovers 98.8% of history-matcher matches, and improves final Average@4 from 0.756 to 0.865; replay audits confirm 100% (781/781) of structured-state matches are executable.
• Fixed-state teacher interventions show SMRC-SD shifts candidate-vs-sampled action log-prob margin by +1.424 vs. no privilege (matched FullPath-SD: +1.158), and the matched-vs-unmatched (A vs C) contrast under FullPath-SD is +0.070 [+0.012, +0.153], evidencing that reference compatibility, not generic task text, drives the effect.

Why it matters / caveats: SMRC-SD shows that privileged distillation for multi-turn agents needs a state-compatibility check before applying reference-conditioned supervision, and that both routing (withholding SDL on unmatched turns) and contextualization (localizing teacher guidance to the reached state) independently and jointly contribute to policy gains; it also yields shorter, less repetitive responses (mean length 78.8 tokens vs. GRPO's 79.5, versus 142.6 for FullPath-SD and 255.6 for Skill-SD) — the method relies on hand-engineered, environment-specific state signatures/adapters, and all privileged components (references, signatures, matching, teacher context) are used only at training time, not at inference.

Problem: It is unknown whether the 70B-parameter scale used by Centaur (an LLM fine-tuned on human behavioural data to serve as a general-purpose cognitive proxy) is actually necessary, and whether such models process genuine task structure or merely exploit statistical shortcuts in the prompt.

Method: The authors fine-tune fourteen base (non-instruct) models from 135M to 14B parameters across four architecture families (Qwen3-Base, Llama-3.1/3.2, SmolLM2/3, OLMo-2/3) on Psych-101 (10.7M trial-level choices from 160 experiments), using rank-stabilised LoRA applied to all linear layers with loss masked to human-response tokens only, following the Centaur recipe. They independently sweep LoRA rank (4–64), training-set size (nested stratified subsets), and model scale, naming the resulting model families Qwentaur, Llama-Centaur, Smoltaur, and Olmotaur. Two diagnostics are then run: (1) structural ablation that progressively strips four prompt channels — instruction, stimuli, feedback, choice history — across 27/32 experiments, including a novel "content-masked" condition that preserves template but replaces stimulus/feedback content with generic placeholders; (2) order-permutation tests on an exchangeable task (THINGS odd-one-out) versus an adaptive/non-exchangeable task (intertemporal choice) as a negative control.

Key results:

• In-distribution (Psych-101, mean NLL over 38 tasks with published cognitive baselines): fine-tuning gain averages 0.188 nats (range 0.138–0.295) across 8 matched models, versus only 0.019–0.024 nats gained from scaling Qwentaur 0.6B→14B or Llama-Centaur 1B→8B — i.e., fine-tuning matters roughly 10x more than scale.
• At LoRA rank 8, Qwentaur-8B and Llama-Centaur-8B (both 0.517 NLL), Qwentaur-4B (0.520), and Smoltaur-3B (0.523) match or beat the reproduced Centaur-70B (0.524).
• Adapter rank substitutes for scale: Qwentaur-0.6B at r=64 (0.513) matches Qwentaur-14B at r=4 (0.516) with 23x fewer parameters; the size needed to match Centaur-70B falls from 8B (r=4) to 0.6B (r≥32).
• Out-of-distribution (Psych-201-RT, 18 held-out experiments): the 8 matched models span 0.788–1.033 NLL (a 0.244 spread) versus only 0.028 in-distribution spread (0.509–0.537) — scale advantage reappears and does not saturate; only 3 of 14 fine-tuned models beat the base Qwen3-14B (0.869) OOD, versus all 14 beating the best base model (0.647) in-distribution.
• Structural ablation: information retention R falls from 1.00 (original) to 0.84 (instruction-ablated) to −0.12 (content-masked) to −0.26 (history-only) (Friedman χ²(3)=63.76, p=9.3×10⁻¹⁴); masking stimulus/feedback content destroys 75.7% of learned information (instruction: 12.5%, template: 11.7%), pushing models below chance — choice history alone does not explain performance.
• Order-permutation: fine-tuned models show near-zero response variance under permutation on the exchangeable THINGS odd-one-out task, but are order-sensitive on the non-exchangeable intertemporal-choice task, matching the tasks' true structure.

Why it matters / caveats: The authors argue small (0.6–1B parameter) cognitively fine-tuned models can serve as practical, cheaper "noise ceiling" estimators for psychological experiments within trained paradigms, since predictions are grounded in actual stimulus/feedback content rather than shortcuts. They caution these models are not mechanistic theories of cognition (they posit no architecture and explain no process), their validated scope is bounded by the paradigms in the training data (Psych-101), and it is untested whether they extrapolate to more extreme conditions; scale still matters for generalization beyond the trained paradigm.

Problem: The standard metric for judging whether a "contamination mitigation" strategy has restored an LLM's genuine capability after benchmark leakage — the Gap of Aggregate Performance (G-AP) — is flawed because it uses discrete 0/1 correctness marks and averages before differencing, letting over- and under-suppression cancel out and masking whether restoration actually occurred per question.

Method: The authors propose SA-PPG (Stratified Aggregate of Per-question Probability Gaps): estimate each question's solve probability by sampling m responses, take the per-question absolute difference (Per-question Probability Gap, PPG) against the clean model, then group questions into B equal-width bins by the clean model's solve probability and average within-group before averaging across groups, which prevents strategies from trivially minimizing the metric by pushing solve probabilities toward the clean model's most frequent (near-zero) values. They also propose RailCap, a decoding-time mitigation strategy: it indexes the contaminated model's own greedy trajectory as n-grams, and whenever a sampled response's last n tokens match a trajectory window (a "fall-back" onto memorized greedy decoding), it caps the logit of the next trajectory token to the level of the current runner-up token, applying this step-wise during generation rather than relying on a one-shot pre-hoc contamination estimate.

Key results:

• On Llama-2/GSM8K, under G-AP, LNE-blocking appears near-perfect (0.0235 vs. 0.3192 for no mitigation) and ranks best; under SA-PPG, LNE-blocking's score reverses to 0.2932 (barely better than Identity's 0.3261), while RailCap — not best under G-AP (0.0728) — becomes best under SA-PPG (0.1914).
• Two independent sampling batches from the same clean model show mean per-question gap of 0.190 under discrete (0/1, m=1) readout vs. 0.041 under solve-probability (m=50) readout — the probabilistic readout is far more reproducible.
• Cancellation effect: LNE-blocking's over-suppression component (0.0836) is 2.0x RailCap's (0.0420), yet its G-APP score (0.0207) reads 3.8x better than RailCap's (0.0794); switching only the aggregation method (G-APP→SA-PPG) moves LNE-blocking's score by a factor of 14 (0.0207 vs. 0.2932).
• A synthetic "All-Zero" trivial strategy (forces solve probability to 0 for every question) scores 0.2190 under equal-weighted A-PPG — beating real strategies like Identity (0.3793), TED (0.3536), and Shortcut (0.2399) — but is correctly ranked worst (0.4903) under SA-PPG.
• Across 6 settings (3 models: Llama-2, Gemma, Pythia × 2 domains: GSM8K, PQ), RailCap attains the lowest SA-PPG in every setting, with its largest margin on Llama-2×GSM8K (0.1914 vs. runner-up Shortcut's 0.2476); TED is nearly indistinguishable from no mitigation (e.g., 0.3250 vs. Identity's 0.3261); LNE-blocking underperforms Identity on all three models in the PQ (paraphrased) domain.
• Ablation on RailCap's n-gram threshold: n=4 gives best SA-PPG (0.1914), balancing residual contamination (0.0942) and collateral damage (0.0972); replacing the runner-up cap with a hard ban raises SA-PPG to 0.2190.

Why it matters / caveats: The paper shows prior contamination-mitigation strategies' claimed "restoration" is substantially overestimated by the widely-used G-AP metric, and provides SA-PPG plus RailCap as more reliable alternatives; results are demonstrated only on GSM8K and a paraphrased variant (PQ) with simulated contamination via LoRA fine-tuning on three specific open model families (Llama-2-7B, Gemma-4-E2B, Pythia-12B), so generality to other benchmarks/models is not directly established in the text provided.

Problem: It is unknown whether personality-conditioned LLM agents (PC-Agents) undergo personality changes after major life events that are systematic, correctly directed, and appropriately scaled/varied relative to documented human personality-psychology patterns — a gap relevant to building coherent "lifelong" agents.

Method: The authors build a four-stage measurement pipeline (baseline BFI-44 Big Five inventory → life-event exposure → first-person reflection → post-event BFI-44) applied to 100 demographically controlled personas (2 genders × 5 cultural regions × 10 Big-Five-anchored personality archetypes) across 11 life events (occupational, social, health) and 11 LLMs, yielding ~12,100 event-conditioned trajectories. They define four diagnostic axes (existence, direction/magnitude, demographic shape, individual shape) anchored against human longitudinal psychology priors (e.g., Specht 2017), and introduce item-level reliability metrics (weighted Cohen's κ, Directional Consistency Ratio DCR) and a composite benchmark, BFI-Adapt, used to rank 14 models. They also run a validation suite (no-event retest, paraphrase robustness, scenario-based decision convergence, delayed retention after unrelated dialogue).

Key results:

• Movement is broad but weakly targeted: within-model median difference in movement rate between event-trait pairs with vs. without documented human directions stays below 0.05.
• Directional agreement (DC%pair) spans 48.1%–70.4% across 11 models (Doubao-Seed-2.0-Pro lowest, Kimi-K2-0905 highest); among the 27 definite-direction event-trait pairs, only 14 (51.9%) matched expected human direction.
• Magnitude calibration: only 11.0%–16.4% of responses fall within the human reference effect-size band (Gemini-3-flash best at 16.4%); the rest are reversed (20.8–40.2%), under-shifted (15.1–54.0%), or overshoot (9.9–31.6%).
• Retirement's documented Conscientiousness decline is universally reversed by every model (median 11.5% correct direction).
• Demographic (gender/cultural-region) moderation is largely absent: median across-strata SD is 0.044 BFI units, and no gender or region comparison survives FDR correction.
• Individual/persona-level dispersion is compressed 3-4x vs. human samples: median σ_LLM = 0.19 vs. human envelope of 0.5-0.8; baseline persona differentiation SD (~0.7) is about 3.6x the event-induced change SD.
• BFI-Adapt leaderboard (14 models) spans 0.348 (Gemini-3-flash, top) to 0.044 (InternLM3-8B-Instruct, bottom), a 7.9x range overall (4.9x among the 11 API models); top three are Gemini-3-flash, GLM-4.6, Qwen3-235B.
• Validation: event-conditioned changes exceed no-event retest noise by 1.6x-9.0x; paraphrase agreement 80.0-92.7%; scenario-decision correlation ρ = 0.003-0.105; short-range retention after unrelated dialogue ranges 62.6%-85.3%.

Why it matters / caveats: Current PC-Agents "simulate the mean of human personality dynamics, but not its shape" — useful for role-playing/companion applications but not yet reliable for psychologically faithful long-term persona evolution. Authors note their pre-post design cannot capture the acute-phase-to-adaptation trajectory documented over years in human panels, and the human-prior direction table (from Specht 2017) is a coarse meta-analytic reference.

Problem: Existing privacy benchmarks for LLM-based agents evaluate only what the agent's final response or outgoing actions disclose, overlooking the earlier "acquisition stage" where agents autonomously pull sensitive information into their context beyond what a task actually requires, leaving that excess data one careless action or attack away from an outright leak.

Method: The authors introduce PrivacyPeek, a benchmark of 1,182 cases spanning 7 acquisition behaviors (Normal-Filename Access, Sensitive-Filename Access, Cross-Format Access, Out-Dated Data Access, Excess-Field Access, Forbidden-Content Access, Out-of-Scope Inferential Access) across 16 application domains (healthcare, finance, legal, etc.), generated via a human-guided pipeline combining author-written templates/seeds with GPT-4o expansion and 10 deterministic quality gates. Each case defines a human-reviewed minimum task scope; agents are evaluated with two complementary methods: Acquisition Inspection, which matches the agent's tool-call trajectory/observations against out-of-scope sensitive items (yielding metrics CER, HCER), and Probe Elicitation, a post-task follow-up question testing whether the retained context supports disclosure of a case-specific sensitive target (yielding PLR, HPLR), evaluated using a Smolagents CodeAgent with 8 fixed tools and a GPT-4o judge.

Key results:

• Across 10 LLM-based agents (GPT, Claude, Llama, Qwen families) on 1,182 cases, Content Exposure Rate (CER) ranges from 6.77% (Llama-3.2-3B-Instruct) to 51.95% (Claude-Sonnet-4), and Probe Leakage Rate (PLR) ranges from 16.67% to 57.53%, exceeding CER for every agent.
• Task completion rate (TCR) correlates strongly with both CER (Spearman's ρ=0.818, p<0.01) and PLR (ρ=0.685, p≈0.029) — Claude-Sonnet-4 has the highest TCR (86.46%) and highest CER (51.95%); GPT-5.1 is an exception, with TCR=85.70% but PLR only 25.38% (32.15 points below Claude-Sonnet-4).
• Out-of-Scope Inferential Access (Task 7) has the highest mean PLR (53.29%) across agents; domains like social media, healthcare, legal, and finance show PLR 36–51%, versus under 25% for retail/engineering/logistics.
• A category-aware system-prompt directive reduces CER by 3.73–12.69 percentage points across agents (e.g., GPT-5.1: 20.56%→11.77%; Claude-Sonnet-4.5: 48.22%→36.47%) but leaves more than half the baseline risk unmitigated.
• Embedding an explicit confidentiality marker in sensitive content actually raises PLR on 5+ agents (e.g., +10.56% for Qwen3-4B-Instruct, +9.87% for Claude-Sonnet-4.5) rather than suppressing leakage.

Why it matters / caveats: The findings show acquisition-stage privacy leakage is widespread and correlated with agent capability, meaning more capable/helpful agents tend to acquire and later risk leaking more sensitive data; the authors note prompt-level defenses and in-content confidentiality markers are insufficient mitigations, arguing that acquisition-stage privacy requires external policy controls (e.g., field-restricted APIs, scoped views, pre-tool policies) rather than relying on the agent's own restraint, and that their benign-agent, no-attacker threat model (no prompt injection or compromised tools) likely understates real-world risk.

Problem: Hard prompt compressors that score and select tokens/sentences/chunks independently can retain a "relevant" fragment (e.g., containing the answer) while deleting a dependent fragment needed to interpret it (e.g., a bridge fact), a failure mode the authors formalize as "referential dangling," and it was unclear how prevalent this is or whether it can be fixed within a fixed token budget.

Method: The authors formally define referential dangling for extractive hard compressors (a retained task-relevant sentence whose every minimal sufficient support set is not fully retained) and prove via a proposition that additive/independent-scoring selection objectives provide no guarantee against it even when a complete, budget-feasible alternative exists. They measure dangling prevalence empirically using a directional content-word coverage metric (threshold θ=0.5) on Beaver (a training-free compressor using Qwen3-0.6B embeddings) and five other hard compressors (PartPrompt, Selective-Context, LLMLingua-2, LongLLMLingua, DAC) across HotpotQA, 2WikiMultiHopQA, MuSiQue, and LongBench-v2. They then test two remedies at fixed budget: (1) an oracle "Reselected" condition that reinserts the omitted annotated supporting paragraph while removing non-supporting paragraphs to preserve budget, and (2) an automatic restoration pipeline using a fine-tuned BERT-base classifier that ranks candidate omitted sentences (by relevance to a retained sentence + question) for reinsertion without using ground-truth annotations at inference.

Key results:

• At compression ratio 0.30, Beaver leaves the answer path incomplete (ρ_d) in 34.2% (HotpotQA) to 54.2% (MuSiQue) of bridge examples; complete evidence retention rate (ρ_e) ranges from 27.0% to 61.0%. Human audit found 95% precision (38/40 flagged examples genuinely required the omitted paragraph).
• Across six compressors on a shared 184-example HotpotQA bridge set at ratio 0.30, dangling rates range from 32.1% (Beaver) to 59.8% (LongLLMLingua); PartPrompt (hierarchical syntactic selection) still dangles at 47.8%.
• Every one of 80 documents in LongBench-v2 Single-Document QA contains at least one dangling reference (mean per-document rate 30.5%, range 25-37% across subdomains).
• Dangling is not explained by low relevance scores: under Beaver similarity, mean salience percentile is 92.8% for definition sentences and 84.1% for answer sentences (both high).
• Fixed-budget reselection (reinserting omitted support paragraph) improves Qwen3-8B accuracy by 28.8-34.3 points (p<10⁻⁴) on the three multi-hop datasets, recovering ≥88% of the gap to "Full support" contexts.
• Stronger models still lose accuracy: GPT-5.5 is 8.8 points less accurate on MuSiQue with compressed (Base) vs. Full-support contexts; GLM-5.2 improves 24.2 points with Full support on MuSiQue.
• Automatic restoration (no ground-truth annotations at inference) improves Qwen3-8B accuracy by 4.7 points (p=0.022) on HotpotQA while only changing compression ratio from 0.30 to 0.31; a random-sentence-addition control only gained 2.0 points (not statistically distinguishable at this sample size, p=0.15).

Why it matters / caveats: The results show referential dangling is a systematic, paradigm-level failure across compressor architectures and scoring signals (not a bug in one implementation), implying hard compressors should optimize for referential completeness, not just relevance. The authors note transfer of their automatic restoration method beyond Beaver/HotpotQA to other compressors and annotation-free QA settings remains unresolved, and some candidate-source comparisons (e.g., "All mentions", "Hybrid") did not reach statistical significance.

Problem: Creative capabilities of multimodal LLMs (MLLMs) are hard to evaluate at scale because creative tasks lack explicit targets and reward signals compared to accuracy-oriented tasks, and no existing benchmark specifically measures "receptive creativity" — the ability to recover an intended meaning from non-obvious but meaningful cross-concept relations.

Method: The authors formalize cross-concept understanding as an encoding/decoding pair (a communicator substitutes a source concept via associative "bridge paths" into displaced but imageable cues; a decoder must reverse this) and build C4, a cognition-inspired framework grounded in Chinese four-character idioms (chengyu) because idioms have fixed exact answers. They construct a manually annotated, third-party-reviewed chengyu-oriented cross-concept network (47 target chengyu, 168 anchored slots, 758 deduplicated bridge chains) via two-round double annotation plus adjudication, then use it to synthetically generate image items at four difficulty levels (L1–L4, defined by bridge count and depth) plus 37 human-created web-collected figures, forming the C4-Eval benchmark (221 items, 84 target chengyu) with five task settings (T1 image-only recovery, T2 bridge-hint recovery, T3 4-way candidate recognition, T4 free explanation with JSON, T5 known-answer explanation) evaluated via strict exact-match scoring on ten MLLMs (5 closed API, 5 open/local via vLLM).

Key results:

• Best primary accuracy (T1–T4 average, 884 cases) is GPT-5.5 at 50.7%, followed by Kimi-K2.6 at 48.0%; the best open-source model (GLM-4.1V-9B) reaches only 18.1%, and the weakest (Qwen3.5-9B) reaches 9.2%.
• Candidate constraints (T3) boost accuracy sharply: candidate recognition exceeds the mean of the three open-recovery settings (T1/T2/T4) by 17.3 to 56.0 percentage points across all ten models.
• Bridge hints (T2) and explanation requests (T4) give only modest gains over image-only recovery (T1) — e.g., GPT-5.5: T1=35.7%, T2=40.3%, T4=39.4%, vs. T3=87.3%.
• RQ3: moving from L1 to L2 (adding a second target slot) causes roughly 3x the average accuracy drop of moving from L2 to L3 (deepening one bridge path); additional depth beyond the first multi-step path yields no further measurable difficulty — slot count dominates difficulty over bridge depth.
• 83 of 221 items are never solved by any of the ten models under any open setting; only 6 items remain unsolved even with candidates provided, indicating shared, systematic decoding failures rather than isolated per-model weaknesses.

Why it matters / caveats: The results show current MLLMs' bottleneck is open answer-space search rather than visual perception or relation inference (visual cues are largely correctly recognized, per T3 gains), and case studies show models often consume visible substitute concepts literally instead of decoding them backward through the intended bridge relations — indicating a genuine gap in creative/analogical reasoning, not just recognition; the benchmark is inherently tied to Chinese-language chengyu and cultural allusions, and L1–L4 difficulty labels are explicitly described by the authors as construction parameters rather than a validated intrinsic psychological difficulty scale.

Problem: Content owners lose practical control over visual assets once they enter AI pipelines, and this survey addresses the lack of a unified account of "protective" adversarial techniques (turning adversarial perturbations into owner-side defenses) that have developed independently across five separate research communities.

Method: The paper is a survey that introduces "adversarial attacks for good" as a unified paradigm covering five lifecycle stages/mechanism families: adversarial privacy filters (against facial recognition at sharing time), unlearnable examples (against unauthorized training), proactive generative safeguards (against malicious editing/imitation by generative models), adversarial CAPTCHAs (against automated agent abuse), and provenance/accountability mechanisms (post-circulation attribution). It proposes a shared evaluation framework of three axes — transferability (L1: white-box/gray-box/black-box), adaptability (L2: static/routine/adaptive manipulations survived), and deployment readiness (L3: laboratory/external/operational) — to make claims across the five communities commensurable, and organizes representative works from each family into a three-level taxonomy.

Key results:

• Surveys and categorizes roughly 190+ representative papers (per the reference list, families numbered up to ~193) across the five families, organized in a three-level taxonomy (family → subcategory → representative works) shown in Fig. 2.
• Table 1 shows this is the first survey to cover all 5 families with both adaptive-robustness and deployment-readiness framing (prior surveys cover at most 1-3 families and lack one or both dimensions).
• Publication-count analysis (Fig. 3) shows explosive growth in all five families in 2023-2026, with generative safeguards and unlearnable examples the largest categories.
• Cross-family analysis finds that most protections in adversarial privacy filters and unlearnable examples are broken by "informed adaptive" countermeasures (e.g., JPEG recompression, blind face restoration, purification, adversarial training) — robustness is largely demonstrated only against static or weakly-adaptive adversaries.

Why it matters / caveats: The authors explicitly note that evidence for these protective methods "beyond controlled benchmarks remains scarce," most techniques are validated only in laboratory settings against static adversaries, and open problems remain around standardizing evaluation, extending protection to multimodal/agentic pipelines (e.g., MLLM visual agents), and defending against adaptive/informed adversaries rather than just naive processing pipelines.

Problem: Existing embodied vision-language models are trained on isolated, task-specific objectives, leaving unclear how the recurring perception/reasoning capabilities a robot needs throughout iterative execution (as the scene continually changes) should be organized, integrated, and verified within a single model.

Method: The authors define an execution-centric capability taxonomy with four families — Spatial Reasoning, Temporal Understanding, Action Guidance, and State Verification — and build Capek 0.5 (2B dense and 35B-A3B MoE, on Qwen3.5-2B/Qwen3.6-35B-A3B backbones) using a two-stage post-training recipe: (1) train four same-origin capability specialists independently via token-level GRPO with capability-specific reward functions (format + accuracy, tIoU, box IoU, Fréchet-distance trajectory reward, progress/next-action reward, etc.), then (2) consolidate the specialists into one inference-time checkpoint via TIES weight-space merging (trimming 80% lowest-magnitude parameters, sign-consensus averaging) followed by routed Multi-Teacher On-Policy Distillation (MOPD) that distills each specialist's behavior onto student-generated rollouts via token-level reverse KL. They also introduce Capek-StateBench, a new benchmark for physical- and task-state verification.

Key results:

• Against protocol-matched Qwen initializations, Capek 0.5-35B-A3B improves 28 of 34 matched benchmark rows, and Capek 0.5-2B improves 30 of 34 matched rows.
• Capek 0.5-35B-A3B shows balanced average performance across capability domains (e.g., ~70.7% Spatial, ~62.7% Temporal, ~68.1% Action Guidance, ~61.5% State Verification, ~82.9% General, per Figure 1), reported as competitive with or ahead of RynnBrain-30B-A3B, MiMo-Embodied-7B, HY-Embodied-1.0-30B-A3B, and Embodied-R1.5-8B on matched active-compute comparisons.
• All benchmark scores are produced via the authors' own "DeepInsight" evaluation harness with reruns for both Capek 0.5 and baselines (no scores quoted from official reports), for consistency.
• The paper reports the consolidated model "retains all four specialized capabilities in one checkpoint with quantified losses" relative to the standalone specialists (specific per-capability retention numbers are in Section 5.3, not fully captured in the extracted excerpt).

Why it matters / caveats: The specialize-then-consolidate recipe (TIES + routed MOPD) offers a practical way to fold heterogeneous, format-incompatible embodied capabilities (points, boxes, trajectories, temporal spans, structured verdicts) into a single deployable autoregressive model without needing task-specific decoding heads; a caveat is that consolidation is not lossless — the paper explicitly measures and reports capability-retention losses after merging rather than claiming perfect transfer, and benchmark comparisons rely on the authors' own re-run evaluation harness rather than published baseline numbers.

Problem: Prior multimodal LLM approaches to affective computing focus on task-specific specialists, neglecting synergy across emotion perception, understanding, and interaction tasks and leaving the models' latent reasoning potential underexplored.

Method: The authors build OneEmo, trained on a new dataset EmoWorld-130K (130K theory-driven reasoning trajectories across 8 tasks: MSA, B-MER, OV-MER, MIR, MHU, MSU, ERG, ESC), constructed via a three-stage pipeline — trajectory distillation from Seed-2.0-Lite guided by psychological theories (Ekman's theory, Incongruity-Resolution Theory, Appraisal Theory, DSM/ICD-11, CBT/ACT), automated reverse verification, and AI-assisted human expert re-annotation. Training uses Emo-Chord, a multi-task RL strategy built on Qwen3.5-4B: two-stage curriculum SFT followed by GRPO with an auxiliary SFT loss (annealed weight μ from 0.5 to 0.02) to prevent policy collapse, a composite reward (format + thought-coherence/factuality + answer accuracy/similarity), and a Task-Aware Linear Decay Gating mechanism to regulate reasoning length.

Key results:

• Perception/understanding (Table II): OneEmo (+Emo-Chord, 4.5B) achieves average scores of 81.13 (MSA), 68.61 (B-MER), 71.22 (OV-MER); outperforms all open-source generalists including 16B Cosmos3-Nano (61.08 avg on B-MER-analog); on MIR (20/30-class intent), OneEmo scores 64.47 avg, beating Gemini-3.1-Pro's 61.39 by 3.08 points; trails MSA leader by only 0.24 points despite ~70x fewer parameters than some commercial models.
• Interaction (automatic eval, Table III): OneEmo(+Emo-Chord) scores Emp. 4.25/Coh. 4.16 on ERG and Ove. 4.67 on ESC, outperforming open-source generalists like Qwen3.5-9B (Ove. 4.51).
• Human evaluation: OneEmo achieves up to 84.33% win rate for "human-likeness" against baselines; wins 71.33%/3.67%/25.00% (win/tie/lose) on ESC overall-effect vs. Qwen3.5-9B; statistical tie against MiMo-v2.5 (310B) on some dimensions.
• Ablation on reward functions: removing the answer reward drops average performance by 2.97 points across 8 tasks; removing thought/format rewards drops it by 1.16/2.7 points respectively.
• Training-strategy ablation: Emo-Chord (RL-v4, curriculum-SFT-stage-1 init + RL+SFT mix) achieves the highest aggregate gains versus alternatives like direct RL+SFT without cold-start (RL-v3), which degrades badly on MIR/MHU.

Why it matters / caveats: OneEmo demonstrates that unifying perception, understanding, and interaction via multi-task RL yields mutual benefits and lets a compact 4.5B model rival or beat much larger commercial models (e.g., GPT-5-mini, MiMo-v2.5) on several affective tasks. The authors note a caveat: after RL, OneEmo shows a marginal decline in ERG informativeness/coherence, attributed to human-annotated references in EmoWorld-130K lacking the exhaustive multi-solution style of commercial chatbot outputs, and top-tier commercial models still retain an edge on interaction-level tasks due to their scale.

Problem: Foundation models for medical imaging (e.g., RETFound for retinal fundus photography) achieve strong performance but are not inherently interpretable, relying on post-hoc attribution methods that are often unfaithful to the model's actual decision-making process — a critical shortcoming in high-stakes clinical domains.

Method: The authors propose Dual-IFM, an interpretable-by-design foundation model using a BagNet-33 backbone (small 33x33-pixel receptive fields that produce faithful class evidence maps) trained with the t-SimCNE self-supervised contrastive algorithm (a SimCLR variant with a 2D projection layer and, per their modification, cosine similarity in stage one rather than t-SimCNE's default Euclidean similarity throughout, which they found gives better visualization quality on their data). The 2D projection layer enables direct, parametric visualization of the full representation space (unlike non-parametric t-SNE, which must be refit for new samples), and a separate classification head with a sparsity constraint is added post-pretraining for fine-tuning while producing spatial class-evidence heatmaps.

Key results:

• Pretrained on 802,360 color fundus images (EyePACS, AREDS, UKBiobank combined) using 18.3M parameters (BagNet-33), vs. RETFound's 303.3M parameters (16x more) trained on 1.6M images.
• Linear probing: Dual-IFM improves over task-trained BagNet (ImageNet-initialized) baseline on 4/7 datasets and over RETFound on 6/7 datasets (e.g., EyePACS 0.849 vs RETFound 0.794; AREDS 0.919 vs 0.871).
• Fine-tuning: Dual-IFM performs on par with or above RETFound on several out-of-distribution datasets (e.g., EyePACS 0.821±0.036 vs RETFound 0.767±0.040), though RETFound leads on others (Glaucoma 0.947 vs 0.925; FIVES 0.936 vs 0.898).
• Inference efficiency: Dual-IFM runs in 81ms/batch vs. RETFound's 140ms/batch (1.7x faster), though with slightly lower peak memory (1.12GB vs 1.31GB).
• Local interpretability: on IDRiD, Dual-IFM's class evidence maps achieve lesion-localization precision of 0.674 (top-10 activated patches vs. DR lesions) compared to 0.384 for RETFound's post-hoc LRP heatmaps.
• 2D embedding quality (k-NN AUROC): Dual-IFM's cosine-similarity variant achieves 0.821 average k-NN AUROC (e.g., APTOS 0.881, Glaucoma Fundus 0.883) vs. 0.745 for the Euclidean-similarity t-SimCNE variant; still somewhat below post-hoc 2D PCA on ImageNet (0.851) or SimCLR (0.839) representations, reflecting the cost of the 2D pretraining constraint.

Why it matters / caveats: Dual-IFM shows that inherent interpretability (faithful local class-evidence maps plus a directly visualizable, dataset-level 2D representation space revealing disease-severity structure and potential spurious correlations/borderline cases) can be obtained without sacrificing representation quality relative to a much larger black-box foundation model, and with faster inference — making it attractive for clinical deployment. Authors note limitations: the work is restricted to color fundus photography (CFP) and used only a limited hyperparameter search during fine-tuning (causing occasional performance drops on PAPILA and FIVES from overfitting), and the BagNet architecture has a high memory footprint relative to its parameter count due to its large spatial feature maps.

Problem: Existing audio-visual embedding models capture semantic ("what") correspondence but discard fine-grained temporal information via global pooling, while dedicated synchronization models capture temporal ("when") offsets but lack strong semantic representations — no prior approach yields a single reusable embedding space that captures both.

Method: The paper proposes FATE (Frame-level Audio-visual Temporal Embedding), built on the PE-AV dual-encoder backbone (fine-tuned via LoRA, updating only 2.3% of parameters) but retaining full frame-level token sequences instead of pooling to a single [CLS] vector. Video features are temporally aligned to the audio timeline via nearest-neighbor downsampling (audio as anchor, since video is temporally redundant and audio carries transient cues), and cross-modal similarity is computed as the mean inner product over aligned frame pairs (the diagonal of the pairwise similarity matrix). Training combines a semantic contrastive loss (symmetric InfoNCE across videos, with cross-GPU all-gathered negatives) and a temporal soft-contrastive loss (cross-entropy against a raised-cosine kernel over candidate temporal offsets within a video), trained on ~183k 10-second clips from VGGSound for 5 epochs on 8 A800 GPUs.

Key results:

• Temporal cross-modal retrieval (AVSync-15, Intra-Video): FATE reaches 22.37 V2A R@1 vs. chance-level ~5.9% and PE-AV's 5.33 (pooling gives near-chance performance); FATE beats Synchformer by wide margins (e.g., 60.83 vs 22.35 V2A N@1 on VGG-Sync).
• Inter-Video retrieval (~850 candidates): FATE reaches 14.22 V2A R@1 / 34.13 R@3 on AVSync-15, far above all baselines (best baseline Synchformer 4.44 R@1/13.56 R@3).
• Zero-shot audio-visual event localization (AVE dataset): FATE achieves 48.3% average accuracy (zero-shot, trained only on VGGSound), beating the best supervised baseline DAM (47.8% avg, trained on AVE with ground-truth annotations) and outperforming zero-shot PE-AV (31.8%) by 16.5 points.
• Generation evaluation metric agreement with human MOS: FATE achieves the best average sample-level Spearman correlation (17.24) and model-ranking correlation (44.41), vs. next-best CAVP (14.15 sample-level) and AV-Align (24.99 ranking).
• Ablations: removing frame-level design (pooling) roughly halves performance (e.g., intra-video V2A R@3 55.74→29.11); using Lsem or Ltemp alone drops V2A R@3 to 19.78 and 15.56 respectively vs. 34.13 combined (super-additive effect); soft temporal labels outperform hard binary labels (34.13 vs 28.50 V2A R@3).
• Inference efficiency: FATE takes 10.8s/video vs. Synchformer's 168.0s (~15x slower), while storage grows from 136KB (pooled) to 3.3MB per video.

Why it matters / caveats: FATE provides a single reusable, frame-level embedding usable across retrieval, localization, and generation evaluation without task-specific heads or fine-tuning, and is far more compute-efficient than specialized synchronization models. The authors note two limitations: absolute sample-level correlation with human synchronization judgments remains modest for all metrics (fine-grained generation evaluation is still an open problem), and the model-ranking correlation is computed over only five generation models, so it should be read as indicative rather than conclusive.