AI papers — 2026-08-17

Problem: On-policy distillation for vision-language models depends on an informative teacher–student asymmetry, but conventional sources of that asymmetry (a larger/stronger teacher, or privileged information such as reference answers or ground-truth regions of interest) become unavailable as model capabilities outpace what humans can supply.

Method: The paper introduces Self-Supervised Visual On-Policy Distillation (S2VOPD), which inverts the usual asymmetry: an exponential-moving-average (EMA) teacher observes the original image while the student generates on-policy rollouts conditioned on a stochastically augmented view of the same image, and the teacher's next-token distribution is distilled into the student's trajectories using a generalized Jensen–Shannon divergence (α=0.5, restricted to the teacher's top-k tokens). The paper systematically searches an augmentation space of four families (information reduction, geometric, photometric, occlusion) and finds the best recipe composes downscaling (0.3–0.6×) with additive Gaussian noise (DDPM forward step t=200). Training uses Qwen3.5-4B/9B with batch size 96, n=8 rollouts, learning rate 2e-6, and EMA teacher decay η=0.05.

Key results:

• S2VOPD improves Qwen3.5-4B from 70.68% to 77.44% average accuracy (+6.76%) across six fine-grained perception benchmarks, above Qwen3-VL-Instruct-235B (75.75%) and GPT-5.4 (72.8%), and matching Qwen3.5-397B.
• Recovers 96% of the improvement achieved by methods using privileged information, while using no ground-truth annotations, rewards, or separately trained teacher.
• Symmetric self-distillation (no augmentation) degrades performance to 65.21% vs. the 70.58% base model under the analysis protocol; removing augmentation entirely (w/o Aug ablation) collapses gains back to 70.52% (vs. base 70.68%).
• Freezing the teacher (w/o EMA) still reaches 75.95%, only 0.40% below the full method, showing the constructed view asymmetry — not teacher self-improvement — drives most of the gain.

Why it matters / caveats: The results show that informative teacher–student asymmetry can be constructed "for free" from augmentation alone, without labels or a stronger model, and that S2VOPD improves both perception and math reasoning where privileged-supervision methods often hurt reasoning. A key caveat found empirically is that stronger augmentation is not always better: augmentations like aggressive cropping that remove question-relevant evidence produce large but uninformative teacher–student gaps and hurt performance.

Problem: Realistic AI-generated videos depicting real-world crises (wars, disasters, public-health emergencies) risk causing misinformation and panic, but existing detection benchmarks give limited evidence on how detectors and generators behave specifically in these socially consequential, real-event settings, or whether detectors hold up as content spreads on social media.

Method: The authors construct RA-Bench, a benchmark using real videos as anchors: 1,830 real-video clips (from 675 public source videos across 10 social-risk L1 categories and 44 L2 subcategories) built through a five-stage pipeline (source collection, PySceneDetect/FFmpeg scene splitting, two-round manual review, postprocessing) and paired with 16,056 clips from four open-source and five closed-source image-to-video (I2V) generators, each conditioned on the real clip's first frame plus a Gemini-3.1-Pro-Preview-generated caption. They evaluate three detector families — seven traditional detectors, ten zero-shot multimodal models under Binary/Diagnostic/Rating prompting, and two fine-tuned MLLMs — and additionally build RA-Bench-HumanProof (a human-deceptive subset) via a 5-reviewer human study and RA-Bench-LastMile (a social dissemination simulation).

Key results:

• Traditional detector AUC falls from public-reference values of 67.6–98.6% to source-level means of only 43.9–57.3% on RA-Bench, with the best-ranked detector changing across generators (public-reference vs. RA-Bench Spearman correlation of only 0.26).
• The strongest zero-shot multimodal model, Gemini-3.1-Pro-Preview, reaches 63.4/62.9 BAcc/macro-F1 under Binary prompting overall, but ranges from 54.3 BAcc on Seedance2.0 to 74.5 on LTX.
• Human reviewers correctly identify 68.6% of open-source generated videos but only 52.9% of closed-source ones (falling to 40.7% for Seedance2.0 and 45.1% for Kling).
• On RA-Bench-HumanProof (633 generated videos labeled "Real" by all five reviewers), Gemini Binary/Diagnostic prompts reach only 54.7%/54.5% BAcc and the seven traditional detectors average just 47.5% AUC.
• The Full condition of the social dissemination simulation reduces mean FakeR across five fine-tuned configurations from 46.0% to 1.4%.

Why it matters / caveats: No detector family generalizes consistently across RA-Bench's generation sources, and the videos that most successfully deceive human viewers are also the hardest for current detectors to flag — with social dissemination (e.g., re-encoding/compression during spread) further weakening detection. The authors conclude current detection methods are not yet reliable enough to protect against realistic AI-generated crisis-event videos as generators continue to improve.

Problem: Existing benchmarks for autonomous AI research-and-development agents evaluate only final task scores, which fail to reveal where progress is gained or lost within a long-horizon experimentation loop, or whether agents actually improve as they accumulate experience.

Method: The paper proposes a framework that decomposes agent behavior into three rule-based, deterministically-computed process metrics — Solution Framing (C1, how quickly good directions are found), Execution (C2, whether changes are reliably delivered), and Feedback Control (C3, whether gains are retained and regressions recovered) — plus a meta-capability analysis using counterfactual comparisons to measure intra-task experience reuse (Mintra, re-running from a branch point with vs. without erased context/notes) and inter-task experience reuse (Minter, transferring extracted "lessons.md" files to held-out tasks). They evaluate seven frontier models (Claude-Opus-4.7, GPT-5.5, Gemini-3.1-Pro, GLM-5.2, Kimi-K2.7-Code, DeepSeek-V4-Pro, LongCat-2.0) on 36 AutoLab tasks spanning Model Development, System Optimization, Puzzle & Challenge, and CUDA, with three rollouts per model-task pair (756 rollouts total), and separately compare Claude Code, model-native harnesses, and OpenCode.

Key results:

• The gap between strongest and weakest model is 0.237 on avg@3 but only 0.122 on best@3, indicating models can reach competitive solutions but differ sharply in reliability across runs.
• Claude-Opus-4.7 leads both avg@3 (0.739) and best@3 (0.790) but costs $89.9/task on average, versus $16.5 (GPT-5.5) and $33.0 (GLM-5.2) for comparable performance.
• Only 3 of 252 best-of-three solutions (1.2%) are classified as genuinely novel approaches after manual review; composition-stacking of established techniques accounts for 111/252 (44.0%), and 16 solutions (6.3%) exploit evaluation-specific shortcuts.
• Intra-task experience reuse generally helps (e.g., LongCat-2.0 shows the largest gain, +0.1454), but inter-task transfer effects diverge across models: it raises DeepSeek-V4-Pro's avg@3 by +0.093 while lowering Gemini-3.1-Pro's by −0.017.
• CUDA tasks show the largest cross-model performance gap (0.403 on avg@3), while Puzzle & Challenge shows the smallest (0.150); harness choice changes run-to-run stability (e.g., native/OpenCode harnesses raise Kimi-K2.7-Code's avg@3 by up to +0.055) but not model ordering.

Why it matters / caveats: The findings indicate current agents behave more like engineering optimizers that recombine established techniques than autonomous researchers capable of genuine methodological novelty, and that a single final score conceals distinct process bottlenecks, inconsistent benefits from experience reuse (which can also backfire by anchoring agents to misleading conclusions), and harness-dependent stability. The full evaluation cost approximately $100,000 in model inference.

Problem: Standard Transformer architectures tightly couple knowledge storage (FFN) and reasoning (self-attention) within each layer via unidirectional forward residual connections, forcing models to rely on long, often verbose token-mediated chains of thought to move information backward through layers, which limits data efficiency and inflates inference cost as scaling returns diminish.

Method: The paper proposes Mobius-v0, which decouples FFN knowledge storage from Self-Attention reasoning by replacing per-layer FFNs with a single globally shared knowledge-vector Memory that all Reasoner (Self-Attention) layers can query, implemented via horizontal concatenation of knowledge vectors (with MoE-style block-wise sparse activation at larger scale). This yields two properties the authors call "Backward Residual Connection" (deep layers can access shallow-layer knowledge through the shared memory, not just via forward residuals) and "Dynamic Latent Reasoning" (reasoners iteratively refine hidden-state latents against the memory over just a few layers before decoding multiple tokens, rather than requiring full-layer passes per token). They validate with a training-from-scratch (TFS) 7B-A1B MoE model trained on up to 1TB tokens against a matched Transformer baseline, and a continual-pretraining (CPT) run, Intern-S2-Mobius-35B, initialized from Qwen3.5-35B-A3B and further trained with SFT and RL.

Key results:

• The TFS 7B Mobius model matches the Transformer baseline's MMLU score using only 62.6% of its training data tokens (1.6× data efficiency).
• Intern-S2-Mobius-35B achieves nearly 4× end-to-end inference speedup over the Qwen3.5-35B Transformer baseline while matching or exceeding downstream scores.
• On general benchmarks, Intern-S2-Mobius-35B outperforms Qwen3.5-35B on average score (67.88 vs. 65.05), including MMLU Pro (89.05 vs. 85.31) and AIME 2026 (95.31 vs. 92.08).
• On scientific benchmarks (Biology-Instructions, Mol-Instructions, MolecularIQ), Intern-S2-Mobius-35B reaches an average score of 52.14 vs. 18.20 for Qwen3.5-35B.
• A case study on a linear-algebra MMLU-Pro question shows Mobius reaching the same correct answer using 516 tokens versus 2,364 tokens for the Transformer baseline, attributed to eliminating repeated derivation and checks.

Why it matters / caveats: The results suggest an architectural (rather than purely scaling-based or linear-attention-based) route to jointly improving training data efficiency and inference efficiency, with the authors additionally arguing Mobius's knowledge-reasoning decoupling has potential for self-evolving models, world models, and hardware-software co-design (e.g., storing knowledge on SSD, reasoning parameters in GPU memory). The paper explicitly states that the mechanisms behind Mobius's data efficiency and shorter reasoning chains "remain to be fully explored" and are only hypothesized, and full training-from-scratch validation at large scale (35B) was not performed due to pretraining cost — the 35B results come from continual pretraining on an existing base model instead.

Problem: Interactive game world models that autoregress video directly in pixel or latent space leave pose, geometry, occlusion, and control implicit in the generative sequence, so errors compound over long rollouts and consistency/controllability degrade.

Method: Marionette factors the world model into three components: a two-stage autoregressive dynamics model (a compact 4-layer/width-256 ~2.5M-parameter "ActionGPT" that autoregresses discrete per-entity action tokens, feeding an 8-layer/width-512 ~25M-parameter "PoseGPT" that turns actions into continuous body pose) predicting an explicit, interpretable 276-dimensional world state (root trajectories, root-relative joints, 6D rotations for two articulated entities plus a weapon sub-state); a zero-parameter deterministic "graphics bridge" that reconstructs metric world-space skeletons via Gram-Schmidt orthonormalization of 6D rotations and cumulative-sum root integration, then rasterizes pose-control video with closed-form occlusion; and a control-conditioned video-diffusion observation model (Wan2.2-Fun-5B, a DiT backbone) that renders photorealistic RGB from the pose-control video, using chunk-relay rollout for long horizons.

Key results:

• Forcing a mismatched action stream changes root-aligned joint error by 31% across 48 held-out segments, showing the state responds causally to actions.
• Left unconstrained, two generated characters drift to 21.2 m apart (recorded sessions stay near 5 m) and a third of frames show ground penetration; imposing a terrain collider and separation cap on the state cuts penetration by 66% with no change to the observation model.
• Routing appearance through the predicted state costs no detectable fidelity: FVD of 831 vs. 799 for recorded pose.
• Against an end-to-end pixel-autoregressive baseline sharing the same backbone and training footage, Marionette achieves FVD16 of 831 [963,1209] vs. 975 [1101,1346] for the baseline.

Why it matters / caveats: By delegating exact geometry/occlusion/motion bookkeeping to a zero-parameter deterministic renderer and leaving only appearance to the neural model, long-horizon failures (drift, penetration) can be diagnosed and repaired purely in the explicit state without retraining the observation model. The dynamics model's current scope is limited to a single monster and player character.

Problem: Existing long-term memory benchmarks are inadequate for realistic on-device mobile assistants, where experiences are heterogeneous, multimodal, cross-app, evolving, and deeply personal, unlike the cloud-based, dialogue-only assumptions of prior work.

Method: The paper introduces KEME (Knowledge-guided Experience synthesis for evolving MEmory), a synthesis pipeline that alternates top-down temporal planning and bottom-up experience evolution to build coherent, temporally consistent long-horizon user trajectories grounded in prior knowledge (personas, temporal constraints, existing interactions), plus a question-answer synthesis stage and a quality-control stage. From this it instantiates two benchmarks: MobileMem (text-only, app-to-assistant integration via predefined templates, covering multi-hop, temporal reasoning, knowledge updating, implicit preference inference, and abstention) and MobileMem-Omni (multimodal, screenshot-sharing setting with visual reasoning and bilingual role-based interactions). Evaluated memory systems include Long Context, NaiveRAG, HippoRAG2, A-MEM, LangMem, Mem0/Mem0-graph, MemOS, and EverMemOS, on GPT-4.1-mini and GPT-5.4-mini backbones, judged by an LLM-as-a-Judge protocol.

Key results:

• On MobileMem with GPT-5.4-mini, HippoRAG2 reaches an overall score of 80.06 (vs. 54.35 for Long Context and 58.70 for NaiveRAG).
• On MobileMem with GPT-4.1-mini, A-MEM reaches an overall score of 79.68, the best among the compared systems.
• On MobileMem-Omni, EverMemOS with GPT-5.4-mini achieves an LLM-Judge score of 39.41%, substantially higher than 24.76% with Qwen3-VL-8B-Instruct as backbone.
• The LLM-judge model (Qwen3-14B) achieves 80.7% agreement with human evaluation on a sampled set of 1,320 examples.

Why it matters / caveats: Performance drops sharply from single-hop to multi-hop to query-focused-summarization questions, and most systems perform weakly on temporal-reasoning questions, indicating current memory construction/retrieval mechanisms still struggle with evolving, temporally-dependent personal data. Answer-generation errors (irrelevant-memory distraction, overlooked evidence) remain a shared bottleneck across all evaluated systems regardless of their dominant failure mode.

Problem: Applying on-policy distillation (OPD) to transfer reasoning from long-context teacher models to short-context, possibly different-family student models is hampered by tokenizer mismatch, teacher-student distribution mismatch, response-length explosion, and training instability.

Method: SimpleOPD performs OPD in a shared text space: the student samples a response under its own tokenizer, the teacher re-tokenizes the same response text (using its own chat template and tokenizer), and only teacher/student tokens whose text spans exactly coincide (found via a linear two-pointer alignment scan) receive teacher log-probability supervision, with unmatched positions falling back to the student's own log-probability — the resulting objective is a token-aligned surrogate for reverse KL, optimized with a PPO-clipped policy loss. To curb length explosion and instability, the method adds a student-reference KL-divergence loss against the student's initial policy and masks the advantages of special termination tokens (e.g., </think>, <|im_end|>) so teacher supervision cannot suppress termination. It is applied to distill the long-context reasoning model SU-01 into student models spanning Qwen3, Qwen3.5, Intern-S2-Preview, GLM-4.7, and Gemma-4.

Key results:

• Intern-S2-Preview improves by 21.2 points on ProofBench, from 34.0 to 55.2, surpassing Gemini-2.5-Pro (52.9).
• Intern-S2-OPD improves on HiPhO from 38.6 to 41.1, and shows gains on HLE and FrontierScience-Olympiad, despite training only on mathematical reasoning data.
• With SU-01→GLM-4.7-Flash, ProofBench improves from 30.8 to 39.7 and AnswerBench from 69.6 to 72.0.
• With SU-01→Gemma-4-26B-A4B, ProofBench improves from 25.5 to 34.2 and AnswerBench from 68.8 to 67.5.
• Scaling the student to Qwen3-30B-A3B yields ProofBench and AIME25 improvements of 15.33 and 5.42 points respectively.

Why it matters / caveats: The results show cross-tokenizer, cross-family OPD is achievable without any supervised fine-tuning on teacher trajectories and that transferred reasoning capability generalizes beyond the mathematics training domain to science benchmarks. The paper notes larger teacher-student capability gaps generally require stronger KL regularization, and without the stabilization techniques, direct OPD causes excessive length growth and frequent truncation (case studies show failure modes like hundreds of repeated self-checks or collapse into one-token loops).

Problem: Kinematic error metrics for humanoid motion tracking (e.g., mean joint error) miss physical artifacts like foot skating and mistimed contacts that determine whether tracked motion actually looks stable and human-like, and existing evaluation suites (e.g., a 140-sequence AMASS test set) are too small and undiverse to stress contact-rich, long-horizon behaviors.

Method: The paper introduces the HumanTracker benchmark, ~153 hours of optical motion-capture trajectories from 24 professional performers, retargeted to a humanoid via General Motion Retargeting (GMR) and organized into four motion families (Daily, Highly Dynamic, Interaction, Ground) with text labels, split 9:1 into train/test. It also proposes HumanScore, a preference-aligned reward model trained with a Bradley-Terry loss on 12K human-annotated pairwise preferences (24K motions, using a temporal Transformer encoder with multi-head self-attention over pose/velocity/contact features and a learned reward head) to predict human preference between two tracking rollouts. Standardized evaluation converts every tracker's output to a common 29-DoF qpos representation and executes it through a shared MuJoCo evaluation entry point, reporting Succ (completion rate), MPJPE (mean absolute joint-angle error), and HumanScore.

Key results:

• Zero-shot evaluation of GMT, TWIST2, SONIC, and Humanoid-GPT shows Humanoid-GPT is the strongest overall tracker, e.g., on Daily: Succ 94.4%, MPJPE 0.046 rad, HumanScore 54.7 (vs. GMT's Succ 17.0%, MPJPE 0.250, HumanScore 2.4).
• HumanScore attains an Align Rate of 0.9083 with human preferences, higher than any individual analytic diagnostic (e.g., MPJPE Align Rate 0.8049, foot-contact accuracy 0.7882).
• On the Highly Dynamic family, SONIC and Humanoid-GPT reach Succ of 82.1% and 86.9% respectively, versus GMT's 36.2%.

Why it matters / caveats: Two rollouts can have similar kinematic (MPJPE) error yet differ substantially in perceived stability, and traditional metrics are not always aligned with HumanScore's family-level results (e.g., SONIC wins on Ground HumanScore despite Humanoid-GPT's stronger completion rate elsewhere), showing no single rule-based diagnostic captures full human-judged trajectory quality.

Problem: Current LLM development depends on massive, often non-permissible datasets, which creates a prohibitive barrier for projects (such as the Danish Foundation Models initiative) committed to training exclusively on ethically sourced, openly licensed, or otherwise permissible data.

Method: Mimir v1 is a 1B-parameter model built on the Hierarchical Reasoning Model (HRM-Text) architecture (hidden size 1,536, 12 attention heads, 2 H-cycles/3 L-cycles, RoPE positional embeddings, pre-norm layer normalization), trained from scratch with the Gemma-4 tokenizer using FSDP, bfloat16 compute, and the AdamW optimizer. It is trained on a curated mixture of 161 permissible datasets (70.5B tokens/epoch), and the authors introduce synthetic "transplant datasets" — generated with Gemma4 31B and quality-audited — to replace non-permissible source data (e.g., portions of the Sapient Flan/Platypus/tasksource collection) with compliant equivalents.

Key results:

• Mimir 1B improves 36.7% over the baseline HRM-Text 1B on Math & Code (64.1 vs. 46.9 average).
• Mimir trails Qwen 3.5 4B by only 0.3 points on average English benchmarks, and SmolLM3 3B by only 3.8% on Math & Code.
• Mimir sets new state-of-the-art results among tested models on Danish grammatical tasks (DaLA, GEC) and WikiQA, and leads its weight class on GSM8K and HumanEval.
• Trained for 1.65M steps on 8 NVIDIA B200 GPUs in under 3 weeks (~1.1s/step).

Why it matters / caveats: The work shows permissible-only post-training data can yield frontier-competitive small models for low-resource languages like Danish, lowering the barrier for ethically constrained research groups. The authors note Mimir still lags behind Gemma 4 (5B, effective 2.3B) on Math & Code, and that its assistant capabilities remain limited compared to state-of-the-art models; reinforcement learning for this architecture is left for future work.

Problem: Existing image editing benchmarks are confined to simple, largely saturated single-image tasks and ignore multi-image editing, real-world deployment scenarios, and highly demanding reasoning-based editing, so they cannot reliably differentiate model performance.

Method: CPI-Bench comprises three subsets built via a four-stage pipeline (taxonomy definition, image collection/curation, instruction generation with human-in-the-loop quality assurance, and privacy-preserving face anonymization): CPI-General-Bench (30 tasks, 2,039 samples, including 10 pioneering multi-image editing tasks), CPI-Practical-Bench (51 real-world application scenarios, 558 samples across 4 domains such as portrait enhancement and e-commerce creativity), and CPI-Intelligent-Bench (1,181 reasoning instances across 8 expert domains/67 sub-disciplines, refined from the authors' prior ExpertVerse benchmark). Evaluation uses an automated Vision-Language Model (VLM) framework with task-specific scoring prompts rating Instruction Adherence, Visual Naturalness, and Physical & Detail Consistency (with tailored dimensions for text-editing, style-transfer, and identity-consistent tasks).

Key results:

• CPI-Bench achieves the highest Spearman correlation and lowest Mean Absolute Error against the Arena Image Edit Leaderboard among all evaluated benchmarks, with model rank positions matching Arena exactly except FLUX.2-klein-9B (one rank gap).
• CPI-Overall score variance across models is 0.2876, far higher than GEdit-Bench (0.0634), ImgEdit-Bench (0.0393), and REDEdit-Bench (0.0585), showing much stronger performance differentiation.
• On CPI-General multi-image tasks, only top-tier closed-source models exceed a score of 4.0 while open-source models plateau around 3.0.
• CPI-Intelligent-Bench variance reaches 0.7282, the largest among all subsets, with open-source models (lacking prompt engineering) scoring notably lower than closed-source models.

Why it matters / caveats: CPI-Bench breaks through the "performance saturation" seen on prior single-image benchmarks, exposing that the main bottleneck separating open- from closed-source editing models is multi-image consistency/reasoning and real-world deployment robustness, giving clearer direction for future model optimization.

Problem: Existing test-time scaling reliability signals (token probabilities, entropy, whole-trace self-consistency) fail to capture the decision-critical semantic content that determines answer correctness, since routine tokens dilute the signal and obscure localized but fatal reasoning errors.

Method: The paper proposes Claim-Level Reliability Assessment (CLR), a training-free two-stage framework. In Stage 1, the model samples K solution traces, each appended with M decision-critical claims. In Stage 2, the same model performs falsification-based verification, searching each claim (given only the problem and claim list) for a decisive contradiction, counterexample, or logical/factual error, producing a binary refuted/not-refuted verdict per claim. The fraction of surviving claims sk is converted into a nonlinear trace-level reliability score rk = sk^M, which is used for reliability-weighted aggregation (Eq. 3) in place of plain vote-counting self-consistency (Cons@K).

Key results:

• On GPT-OSS-20B/CMIMC25, CLR@32 raises accuracy from 77.50% (Cons@64) to 82.19% while using 37.0% fewer generated tokens, exceeding pass@1 by 27.15 percentage points.
• On Gemma-4-12B-it/HMMT25, CLR@32 raises accuracy from 76.67% (Cons@64) to 88.75%.
• Pooled across 16 benchmark–budget settings, CLR's rescue rate of incorrect consensus (recovering a correct answer despite majority error) spans ~16–48%, averaging about 37%.
• Increasing the number of claims M from 1 to 3 improves accuracy by 3.13–3.79 percentage points across benchmarks (Table 3, GPT-OSS-20B).

Why it matters / caveats: CLR shows that reallocating test-time compute from additional sampling to targeted falsification can improve accuracy, token efficiency, or both, without any model retraining or separate verifier. The authors note CLR is not uniformly dominant — gains are smaller on the near-saturated Qwen3.5-27B, and on Gemma-4-12B-it improvements come with 22.2–47.8% more tokens rather than savings; the accuracy–efficiency profile depends on the base model's operating regime.

Problem: Binary success rates and manually-defined rule-based scores are insufficient for evaluating embodied robot manipulation models, since they collapse highly variable trajectory quality (both among failures and among successes) into a single coarse number, obscuring where and why models fail or recover.

Method: PRM-as-a-Judge 1.5 converts rollout videos into dense progress curves using process reward models (PRMs) such as Robo-Dopamine, then computes an OPD (Outcome–Process–Diagnosis) metric suite. Building on version 1.0's metrics (Milestone Coverage, Max Progress, Path-weighted Progress Length, Cumulative Regret Area, Stagnation Ratio), it adds three new conditioned diagnosis-level metrics: Failure Near-Success (FNS, a composite of MP/MC@50/MC@75 for failed rollouts), Drawdown Recovery Ratio (DRR, largest recovery over largest drawdown), and Success Quality Score (SQS, a composite of PPL/CRA/STR for successful rollouts). The toolkit also introduces RoboPulse++, an interval-level benchmark for testing whether PRMs can correctly distinguish increasing vs. decreasing task progress, and is evaluated on the RoboDojo-RealWorld and RoboDojo-Sim benchmarks across VLA and WAM embodied models.

Key results:

• VLAs occupy a substantially higher share of Top-3, Top-5, and Top-10 rankings than WAMs on RoboDojo-Sim (e.g., 64% vs. 36% in one comparison band, per Figure 3).
• π0.5 ranks best overall across the 11 evaluation metrics on RoboDojo-Sim (Table 4).
• Simulation-to-real-world rank correlation is weak: Spearman ρ = 0.18–0.58 across Success Rate, Max Progress, and FNS for the 9 shared models.
• Precision tasks are the best-performing task category (MC@25, MP, FNS), while Open-Vocabulary tasks are the most challenging, with tightly clustered low scores.

Why it matters / caveats: The process-level metrics reveal model rankings inconsistent with binary success rate (e.g., SR ranking diverges from SQS/DRR/FNS rankings), and show that larger model size does not guarantee better performance. The authors caution that simulation performance is a poor proxy for real-world performance, especially on tasks requiring precise alignment or complex contact dynamics, and note limitations in current PRMs around temporal context and negative-progress supervision as directions for future work.

Problem: Existing on-policy self-distillation (OPSD) methods rely on designer-specified privileged context (e.g., fixed answers, feedback, skills, or trajectories) for the self-teacher, which limits the end-to-end learnability and scalability needed for continual agent self-improvement.

Method: Latent On-Policy Self-Distillation (LOPD) makes the teacher's privileged context itself learnable: a dense retriever pulls the top-J relevant prior experiences from an experience bank, and a composer (a LoRA-tuned encoder plus a QFormer-style cross-attention compressor with learned queries) turns them into K continuous latent tokens per experience that condition a frozen-backbone self-teacher. The student rolls out trajectories on-policy and is trained via reverse-KL distillation against the teacher's top-M-plus-tail token distribution at every visited prefix. A privileged-margin constraint (a dual-variable Lagrangian penalty tied to outcome-weighted teacher log-probability advantage) prevents the composer from collapsing the teacher toward the student.

Key results:

• LOPD achieves the best aggregate result in all ten backbone–benchmark comparisons tested (tool-use and coding, across QWEN3-4B, QWEN3-8B, OLMO3-7B).
• On QWEN3-8B: EnvScaler 66.4 vs. 60.2 for the strongest baseline, BFCL-v3 29.88 vs. 29.00, ACEBench 62.7 vs. 58.0.
• On coding: LiveCodeBench/EvalPlus aggregates reach 48.78/81.36 with QWEN3-4B and 50.98/78.41 with OLMO3-7B, beating the strongest alternatives by 2.69 and 0.55 points.
• LOPD surpasses GRPO and Skill-SD with less than 30% of their rollout budget, reaching 0.637 mean reward by generation 576 of a 1,600-generation budget.
• Ablation: the privileged margin is necessary — training with m=0.05 yields 0.637 reward vs. 0.573 for a frozen composer and 0.551 with no margin constraint (m=0).

Why it matters / caveats: The results argue that experience should be treated as a learnable supervision substrate rather than a hand-crafted artifact, since fixed-context OPSD baselines (e.g., SDPO, OPSD) sometimes underperform the vanilla model on certain benchmarks. A caveat noted by the authors: decoded latent tokens remain fragmented, distributed representations that do not reproduce readable procedures, so direct interpretability of what the composer encodes is limited.

Problem: Time-series foundation models (TSFMs) applied to hourly US equity returns produce "forecast collapse" — near-flat, poorly cross-sectionally ranked predictions — while the same pipeline does not collapse when forecasting trading volume on the same stock panel.

Method: The paper derives two theoretical mechanisms for collapse: an exact amplitude-correlation identity showing best-scaled forecast amplitude A* equals |corr(ŷ,y)|, which is bounded by a predictability ceiling ICmax = sqrt(SNR/(1+SNR)); and an invariance proposition showing that any "pointwise" per-series risk (MSE, quantile, likelihood, cross-entropy, etc.) is a functional of per-coordinate laws alone and cannot identify cross-series (cross-sectional) correlation structure. Building on this calibration-ranking tradeoff, the authors introduce CalibRank, an objective combining MSE with cross-sectional correlation via a weight λ, and validate it on Finance1K (a released hourly panel of 1,000 US equities with matched return/volume targets), synthetic ground-truth panels, twelve forecasting architectures, and 97 public GIFT-Eval configurations for TimesFM and Chronos.

Key results:

• CalibRank nearly triples cross-sectional correlation on Finance1K while keeping forecast amplitude close to the target, improving correlation for every model tested.
• Directly optimizing cross-sectional correlation (IC) instead of MSE inflates amplitude by more than an order of magnitude — examples in the paper show forecasts 20–26x the target amplitude (e.g., 25.63x, 21.10x, 23.22x for three sample equities), effectively 98–100% "off scale."
• A survey of 19 TSFMs finds 15 disclose a forecasting loss (only 4 squared-error-like), and none of the 18 sourced training objectives score the cross-section jointly.
• A constructed two-series Gaussian example holds MSE risk fixed at exactly 1.40 while E[IC] ranges from 0.371 to 0.032 (an order of magnitude) purely by changing cross-series coupling.

Why it matters / caveats: The findings expose a blind spot in conventional per-series time-series evaluation: per-series accuracy metrics can look fine while the cross-series structure needed for downstream ranking decisions (e.g., in finance) is never measured or optimized. CalibRank rewards only one functional (cross-sectional correlation) of the dependence structure, not a full copula/dependence model, so it does not address richer needs like calibrated joint scenarios or tail dependence.

Problem: In ReAct-style LLM agent loops, deliberate reasoning is confined to the Thought phase; the recurring Action–Observation interval, while the agent serializes an action and waits for the environment, is a "reasoning idle window" during which no additional reasoning takes place.

Method: Second Thought is a training-free inference framework that, the instant a Thought phase concludes, forks four auxiliary branches — Check, Recall, Rehearse, and Alternative — that continue the same conversation snapshot and decode concurrently with the main thread's action serialization and tool execution (sharing the prefix KV cache). Each branch streams output as interruption-safe "atomic thoughts" (self-contained, ≤25-word units wrapped in XML tags); when the environment observation arrives, branches are cancelled and up to 5 completed atomic thoughts per dimension are harvested and appended to the tool-observation message for the next turn. It is compared against an unmodified ReAct baseline and against "s1," a compute-matched budget-forcing control that places the same extra reasoning tokens on the main thread's critical path.

Key results:

• Lowers average turn count in all nine model–benchmark pairs (three LLMs × three benchmarks: SWE-Bench Pro, Terminal-Bench 2.1, τ3-bench) and reduces main-thread decoding in six of them by up to 43% (36,519 → 20,798 tokens on SWE-Bench Pro with Qwen3.6-Plus), roughly 20% on average among those settings.
• Pass@1 is statistically unchanged in seven of nine pairs; the two significant changes are gains of +12.4 and +10.2 points (both on Terminal-Bench 2.1).
• A paired wall-clock replay on 50 SWE-Bench Pro instances shows median per-task latency drops from 256.9s to 229.0s (−10.9%).
• Against the compute-matched s1 control, Second Thought attains strictly higher Pass@1 with 1.3× to 3.2× less sequential decoding in all four settings where s1 applies.
• Running all four branches increases per-task API cost by 66.4% to 181.5% across models, driven mainly by cached prefix reads rather than output generation.

Why it matters / caveats: The method relocates added reasoning off the critical path without any training, but benefits are domain-dependent — gains are smaller on τ3-bench (banking), which the authors attribute to short idle windows and failures driven by retrieval/policy adherence rather than planning errors that the four branches target. The added API cost from running extra branches is a stated tradeoff, mitigated but not eliminated by keeping only the top-performing branch.

Problem: The internal features responsible for MLLM visual behaviors are difficult to identify, audit, or control, because sparse autoencoders (SAEs) trained directly on MLLM activations mix features inherited from the language backbone with those altered by multimodal training, and are not directly useful for targeted intervention.

Method: MMDiff trains multimodal SAEs (TopK SAEs for LLaVA-MORE and InternVL3.5-2B, JumpReLU SAEs for PaliGemma 2) warm-started from base-LM SAE dictionaries (LLaMA-Scope, Qwen-Scope, Gemma-Scope) on masked text-only token activations, then diffs each base-LM SAE against its multimodal-adapted counterpart using decoder-direction cosine similarity (rotation) and mean-squared visual activation (visual energy) to isolate an "adapted" feature set. Within this set, per-token contrastive firing analysis (odds ratio and Fisher's exact test between a target and baseline distribution, plus a lexical-invariance filter using neutral prompts) identifies task-specific features for spatial reasoning, multimodal safety, and OCR. These features are then used for causal removal (orthogonal projection of the decoder direction at text-token positions) and MMDiff-CAA steering, which combines multi-layer contrastive activation addition with injection of a discovered feature's decoder direction at its associated layer.

Key results:

• Removing MMDiff-discovered features degrades target behaviors by an average of 12% on spatial tasks (VSR) and 17% on OCR (OCRBench categories), and reduces attack success rate by 24% on multimodal safety attacks (VLSBench), with no impact on VQA performance.
• MMDiff-CAA steering improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer CAA steering baseline (peak gains of +15.4% on a spatial relation and +10.58% on an OCR feature).
• Per-category top unsafe features reduce VLSBench ASR by 17–28% each with |ΔVQA| ≤ 1% and |ΔCtrl| ≤ 1%; across 1,061 candidate safety features the mean effect is ΔASR = −9.67% with ΔVQA = −0.03%.
• Cross-stage ablation on PaliGemma 2 shows instruction tuning amplifies the causal contribution of spatial features by roughly 3× on average versus the pretrained-only checkpoint.
• Ablation confirms the full pipeline is needed: dropping the adapted-feature (model-diffing) filter degrades general VQA by 24–26%, and running the same contrastive-firing selection on a non-diffed, randomly-initialized SAE yields features whose ablation leaves VSR essentially unchanged (+0.22) versus −10.11 for MMDiff features on the same model.

Why it matters / caveats: The results position multimodal SAEs as feature-level interfaces for auditing, localizing, and controlling MLLM behavior, not just for inspection. Stated limitations include: only three backbones were tested, with safety and OCR evaluated on PaliGemma 2 only; MMDiff-CAA requires access to an instruction-tuned reference model; and a minority of safety-related features cause generation collapse rather than refusal when ablated, requiring a post-hoc filter.

Problem: As LLM training-token budgets grow with model size, high-quality domain data (e.g., code, math, Wikipedia, medical text) becomes diluted in the training mixture, and it is unclear how much this scarce data should be repeated to counteract dilution without causing overfitting.

Method: The authors train models of varying sizes at a fixed tokens-per-parameter ratio (TPP), sweeping the fraction of unique high-quality tokens (α ∈ {1/40, 1/20, 1/10}) and repetition count (e ∈ {1,...,7}) for four domains (Code, Math, Wiki, Medical), filling the remaining budget with non-repeated web data, and optimizing with the Muon optimizer. They fit a quadratic function to final validation loss versus repetition count to estimate the optimal repetition count, and support the empirical findings with a theoretical one-hot linear regression model (with source-condition assumptions) analyzed via gradient descent, decomposing risk into knowledge-acquisition and noise-fitting error terms.

Key results:

• The optimal repetition count correlates with minimum validation loss at Pearson r = −0.944, versus only 0.400 for model size and 0.018 for the unique-data fraction.
• Under fixed TPP, Math supports the most repetition (optimal ~5–6 repeats), while Medical has the lowest optimal repetition count (~3–4).
• Delaying the learning-rate decay (or using a constant LR) allows the model to tolerate substantially more repetitions before validation loss degrades, compared to early WSD decay.
• Replacing unique high-quality tokens with repeated tokens (at fixed total high-quality fraction) has only a limited effect on out-of-distribution (ArXiv, News) validation loss, unlike the domain-dependent degradation seen in-domain.

Why it matters / caveats: The findings support a practical recipe: tune repetition counts on smaller proxy models at the same TPP as conservative estimates for larger models. The study only repeats one high-quality domain per run and does not test simultaneous multi-domain repetition, which the authors flag as future work.

Problem: On-policy RL with verifiable rewards (RLVR) can improve a model's current objective while making successful behaviors needed for a later training objective too rare to sample and reinforce, a forward-looking risk distinct from catastrophic forgetting.

Method: The paper defines "effective rewardable support" as reward-positive trajectories reachable within a fixed rollout budget, and studies "verifier-induced support reshaping" through repeated verifier-scored sampling and bidirectional GRPO-based training (Math-RLVR on the MATH split vs. IF-RLVR on IFTrain) on Qwen3-8B-Base and Qwen2.5-Math-7B, including sequential training with the opposite verifier. It uses token-distribution analyses (Jensen–Shannon divergence by response position), controlled opening interventions (forcing DRI vs. DAI response openings), reference-policy KL constraints, a DRI-biased SFT prior, and on-policy distillation (OPD) as mitigation strategies.

Key results:

• On IFEval with Qwen3-8B-Base, Math-RLVR raises pass@1 by 6.5 percentage points while best@32 falls by 9.8 percentage points.
• The first generated token shows the highest mean JS divergence between Base and RLVR policies, with first-token-to-interior JS ratios ranging from 9.8× to 106.7× on AIME.
• After IF-first training, the mixed-outcome prompt group (needed for reward variation) falls from 33.6% to 1.6% by step 20, leaving little signal for subsequent Math-RLVR.
• On MathIF/ReasonIF, Math-RLVR raises correctness by 11.3/5.6 pp but joint correct-and-constraint-following support (J) rises by only 2.4/1.7 pp; IF-RLVR raises constraint-following by 18.1/15.3 pp but J rises by only 5.5/6.2 pp.

Why it matters / caveats: Endpoint improvements on one verifier do not guarantee future trainability or joint capability; tested mitigations (KL constraints, DRI priors, OPD) only partially preserve cross-task support or trade preservation against target-task gains, and results are limited to the tested math/instruction-following pair and two model families.

Problem: Because modern LLMs are trained on heterogeneous web-scale corpora, it is difficult to characterize their prior exposure, making it hard to tell whether methods like in-context learning or post-training produce genuinely new capabilities or merely elicit knowledge already present in pretraining.

Method: The authors build LITTLECURRICULUM, an 88B-token corpus filtered from FineWeb-Edu to contain only U.S. K–5 elementary school material, using a multi-stage pipeline: Age-of-Acquisition (AoA) pre-filtering with Zipf-frequency imputation, LLM-as-a-judge annotation with prompts refined via DSPy/OpenEvolve, a FastText classifier followed by a ModernBERT classifier, rule-based symbolic filtering of mathematical notation, and a frequency-sampling stage removing documents with Beyond-K–5-associated terms. They train LITTLELEARNER, a 5B-parameter Qwen3-architecture model, from scratch on this corpus, and evaluate scaling (0.6B/1.3B/5B), post-training (SFT + GRPO), and in-context learning (few-shot/explanation prompting) as ways to extend capabilities beyond the K–5 boundary.

Key results:

• The filtering pipeline reduces Beyond-K–5 document retention to 0% on the CommonCoreText validation set while retaining ~35% of true K–5 documents.
• On MathCAMPS Grade 8 problems, even at pass@1024, LITTLELEARNER solves fewer than half as many questions as the UNFILTERED control.
• Scaling from 0.6B to 5B improves in-scope (K–5) and near-boundary (Grades 6–7) performance but yields essentially no improvement on Grade 8 (fully out-of-scope) material.
• SFT+GRPO post-training substantially boosts in-scope K–5 accuracy but produces only modest gains Beyond-K–5, with no observed difference between training on K–5-filtered versus Beyond-K–5 data for LITTLELEARNER.

Why it matters / caveats: The results suggest the pretraining filter, not the downstream intervention, sets the model's effective capability ceiling — consistent with the view that pretraining data dominates downstream behavior. The authors caution that LITTLELEARNER should not be interpreted as a model of human child development, and that 5B scale may make some emergent behaviors (e.g., strong ICL) less pronounced than at frontier scales.

Problem: Existing hallucination detectors for large vision-language models (LVLMs) either require expensive full-model fine-tuning, rely on external verifiers that ignore the model's own generation process, or reduce internal signals to isolated features and hand-crafted statistics, discarding spatial, sequential, and relational structure needed for precise token-level localization.

Method: UniProbe builds a directed computational-trace graph over image patches, query tokens, and generated response tokens from a single frozen-LVLM forward pass, with edges derived from attention weights (top-scoring image/query positions retained, causal response-to-response edges). It processes this graph with L alternating blocks, each combining a GNN (message passing over typed attention neighbors) for relational evidence, a ViT for 2-D visual geometry over image nodes, and a (Bi)GRU for sequential response order, followed by a linear head producing per-token hallucination probabilities. A streaming variant (unidirectional GRU) enables hallucination-aware decoding that rejects and resamples flagged tokens, and a self-adaptation strategy fine-tunes the detector on the target model's own CHAIR-labeled captions to address train-test distribution shift.

Key results:

• On MHALO with the GLM-4V backbone, UniProbe reaches F1M/F1IoU of 63.2/52.9 versus 59.1/49.8 for the strongest trained baseline (HaloDet), despite keeping the backbone frozen while HaloDet fine-tunes all weights.
• On POPE object-hallucination detection, UniProbe achieves 63.1 F1 / 90.0 AUC versus 41.0 F1 / 75.0 AUC for the strongest prior detector (Token Grounding).
• During hallucination-aware decoding on GLM-4.1V, self-adapted UniProbe reduces CHAIRi/CHAIRs from 18.0/37.2 (vanilla) to 8.2/16.6, a 55% reduction, at only 1.06× vanilla latency.
• Ablations show removing the relational graph component costs 10.1 F1M and removing response recurrence (GRU) costs 4.6 F1M, and a human preference study shows raters prefer UniProbe's mitigated captions 55% of the time versus 30% for PAS and 15% for HaloDet.

Why it matters / caveats: UniProbe demonstrates that structuring a frozen LVLM's internal computational trace (rather than flattening it into hand-crafted features) yields state-of-the-art detection without touching model weights. The authors note it requires access to hidden states/attention (so it is limited to open-source models or providers with internal access) and needs a separate trained readout per backbone, leaving cross-backbone transfer to future work.

Problem: Scientific figures and tables carry the primary experimental evidence in publications, but digital libraries and vision-language models still struggle to retrieve, extract, and reason over this visual content with the numerical fidelity and domain grounding scientific interpretation requires.

Method: The paper is a forward-looking companion piece to the ALD/E-ImageMiner benchmark and ICDAR 2026 Competition on Information Extraction from Atomic Layer Deposition/Etching Scientific Figures, which uses MinerU to extract figures and text from publications and provides panel-level bounding-box annotations across four tasks: figure classification, data table extraction, summarization, and visual question answering (VQA). The VQA component is explicitly designed using Bloom's revised taxonomy, organizing questions into four domain-grounded families (process-oriented, comparative/trend, structure-property, and application/performance) paired with four answer formats (yes/no, factoid, list, paragraph). The authors propose "scientific conceptual understanding from images" as a long-term benchmark objective and outline a roadmap of extensions (contextual interpretation, hypothesis evaluation, provenance, counterfactual grounding, cross-document synthesis, open-ended agentic research).

Key results:

• The ALD/E-ImageMiner benchmark contains 1,951 figures from 205 publications, organized into 49 figure categories.
• The dataset spans experimental and simulation-based atomic layer deposition/etching literature with expert annotation across the four described tasks.

Why it matters / caveats: This is a vision/perspective paper rather than a systems evaluation paper — it reports no model benchmark scores of its own, instead using the ALD/E-ImageMiner competition dataset (described fully in a companion paper) as a case study to argue for a broader, incrementally expandable multimodal scientific AI research agenda covering more domains, figure types, and reasoning depths.

Problem: The Muon optimizer's cubic-time Newton-Schulz orthogonalization step, compounded by communication overhead when weights are sharded across GPUs, creates a substantial computational bottleneck that limits Muon's practicality at scale.

Method: Dion3 attacks this overhead at every level of the stack with four combined contributions: (1) Gram Newton-Schulz, a mathematically equivalent reformulation that iterates on the small symmetric Gram matrix XX^T instead of the large rectangular matrix, stabilized via a "restart" strategy to counter spurious negative eigenvalues from half-precision rounding; (2) custom CuteDSL GPU kernels for symmetric matrix multiplication (AB and αAB+βC) that exploit Hopper/Blackwell architectures by computing only the lower triangle and copying to the upper triangle; (3) the Dion3 update rule, which subselects a fraction f of rows/columns of the momentum matrix (by largest ℓ1 norm) before orthogonalizing, combined with an error-feedback mechanism adapted from Dion; and (4) megabatching, which groups same-shaped weight matrices into a single all-to-all communication round to reduce communication rounds to O(1) per optimizer step, independent of model depth.

Key results:

• Combined contributions reduce Muon's optimizer step-time overhead from 26x AdamW down to 4x AdamW for a 7B-parameter model on four GH200s (Figure 1).
• Dion3 with row-selection fraction f=1/4 improves validation loss over tuned NorMuon at every model scale tested (3B–14B parameters, ClimbMix dataset), with the largest gap of −0.027 loss at 14B parameters, and downstream accuracy gains of up to +1.0 percentage points (3B) and +0.7 points (14B) across a 12-benchmark macro-average.
• Symmetric kernels plus Gram Newton-Schulz give a combined 1.5x+ speedup over standard Muon; adding fractional updates (f=1/2, f=1/4) yields further 2x and 3.7x reductions, for overall speedups up to 6x (some architectures) or 6.5x depending on configuration.
• Megabatched communication reduces per-GPU optimizer step time by 35% for a 1B model sharded across 8 GPUs (from 80.7ms to 52.1ms).
• On architectures with higher aspect-ratio weight matrices (α=8, e.g., Gemma, MoEs), Gram Newton-Schulz and symmetric kernels alone achieve ~2x speedup.

Why it matters / caveats: Dion3 is released as a drop-in replacement for Muon via the open-source dion package, targeting practitioners who want Muon's optimization benefits without its scaling costs across varied architectures, cluster sizes, and parallelism strategies; the authors note optimizer step time can account for 1%–17% of total LLM training time, so these speedups translate to meaningful end-to-end savings, though benefits are most pronounced for models/settings that are compute- or communication-bound in the optimizer step (e.g., smaller models with many shards, or matrices with high aspect ratios).

Problem: Text classifiers trained on crowdsourced corpora often exploit spurious surface patterns correlated with labels, but existing methods either require manually specifying the feature vocabulary or only partially automate discovery, leaving a gap between dataset-level statistical correlation and actual model-level exploitation unaddressed.

Method: UNMASK is a four-stage, fully automated pipeline requiring no human annotation. Stage 1 uses an LLM (SCGEN_LLM) to generate candidate spurious patterns as executable boolean expressions over unlabeled text, followed by description-based and coverage-based deduplication. Stage 2 (Spurious Feature Validation) applies Fisher's exact test with Benjamini-Hochberg false discovery rate control across a discovery/validation split to statistically confirm label associations. Stage 3 establishes causal model dependence: a spurious-reliance screening step (two-proportion z-test) filters candidates, then a GENERATOR_LLM (GPT-4o) produces minimal counterfactual edits that remove the feature while an independent EVALUATOR_LLM (Qwen3-32B) verifies semantic preservation, with causal exploitation confirmed via paired t-tests on the resulting probability shift (Δp). Stage 4 uses the causally-confirmed boolean features directly as annotation-free group labels for debiasing methods including Deep Feature Reweighting (DFR), Product-of-Experts (PoE), SCER, JTT, and LEACE.

Key results:

• On MNLI, causal verification confirms 9 of 10 discovered spurious features on BERT and 6 of 10 on RoBERTa, consistent across 3 seeds, including a divergence where RoBERTa is immune to three contradiction-class features (effect sizes below ε=0.03) that BERT reliably exploits.
• PoE-IPW-Group improves HANS accuracy by up to 12.58 percentage points over ERM on BERT (64.99% vs. 52.41%), while RoBERTa gains +4.08 pp on HANS.
• On CivilComments-WILDS, annotation-free DFR reaches 71.84% (BERT) and 72.12% (RoBERTa) worst-group accuracy, matching or exceeding the 70.1% reported by Kirichenko et al. (2023) for hand-labeled DFR.
• UNMASK's discovery stage beats the strongest baseline (PMI+SFV) by 3.65 pp worst-group accuracy under a fixed BERT DFR recipe, and recovers 6 of 8 canonical WILDS demographic identity axes with zero demographic labels, versus PMI's zero.
• On sentiment classification (SST-2, IMDB), causal verification correctly returns 0 of 5 and 0 of 8 candidate features as exploited (mean Δp ≈ −0.0026 and −0.0067, not significant), confirming the causal gate does not over-flag features when they are the legitimate task signal.
• A human audit of 100 accepted counterfactuals shows 94/100 gold-label preservation; the EVALUATOR_LLM (Qwen3-32B) reaches 80.3% agreement with human-majority vote on ChaosNLI.

Why it matters / caveats: The authors note UNMASK's design is bounded by its representation: only shortcuts expressible as deterministic boolean predicates can be discovered, so latent/semantic shortcuts, topic/style bias, and distributional artifacts like class imbalance are invisible to the pipeline, and the causal counterfactual-edit stage cannot hold every other input property fixed, meaning Δp only upper-bounds each feature's individual causal contribution when multiple features co-fire.

Problem: Existing approaches address 3D reconstruction, dense correspondence, and spatial reasoning separately with task-specific architectures or external geometric modules, limiting knowledge transfer among complementary representations of the same physical scene.

Method: SPARGen is built on the Bagel Mixture-of-Transformer-Experts (MoT) backbone and casts 3D reconstruction, dense correspondence, and spatial reasoning as instruction-conditioned generation tasks. Dense spatial fields (depth maps, point maps, optical flow) are encoded as image-aligned representations and generated via the model's native rectified-flow (VAE latent space) pathway, while compact structured outputs (camera poses, quantized as quaternion/direction/magnitude tokens) and textual answers are serialized as token sequences and generated autoregressively. Training jointly optimizes a sequence cross-entropy loss (Lseq) and a rectified-flow matching loss (Lfield) over data spanning spatial reasoning (MindCube, OmniSpatial, OST-Bench, SPAR-7M, LLaVA-OneVision), visual geometry (ASE, BlendedMVS, CO3D, ScanNet, etc., with MoGe pseudo-labels), and optical flow (TartanAir, FlyingThings3D, Spring, etc.), for 100K iterations on 64 H100 GPUs. Inference also uses a predict-warp-predict optical flow refinement procedure.

Key results:

• Improves depth AbsRel/δ1 on Sintel and NYU-v2, and reconstruction Acc./Comp. on 7Scenes and camera pose metrics on CO3D v2 (RRA@30=74.32, RTA@30=94.33) versus the unified baseline G2VLM.
• Achieves the highest average score on all four spatial-reasoning benchmarks, improving over the strongest competing non-proprietary result by 9.85 points on MindCube, 1.97 on OmniSpatial, 4.99 on OST, and 24.71 on SPAR.
• Zero-shot optical flow on KITTI: EPE of 4.09 and F1-all of 13.34, outperforming RAFT, GMFlow, and FlowFormer.
• Ablations show removing geometry supervision degrades both optical flow and spatial reasoning, and removing optical flow supervision increases 7Scenes reconstruction error and lowers average SPAR score.

Why it matters / caveats: SPARGen shows spatial perception and reasoning can be unified in a single native multimodal generative model without task-specific regression heads, though the authors note that reliance on a frozen VAE creates a spatial-compression bottleneck for geometric edges and high-precision physical quantities, and like VGGT/G2VLM it cannot recover metric scale.

Problem: Text-only next-token-prediction training in MLLMs provides only indirect supervision for visual perception, leaving fine-grained visual detail underutilized, while existing unified generation-understanding models add inference-time cost and often fail to translate generation training into understanding gains.

Method: GAS (Generation as Auxiliary Supervision) uses Next Embedding Prediction (NEP), a cross-modal generation paradigm that autoregressively predicts continuous target image embeddings (extracted via the same frozen ViT and an EMA-stabilized target projector) directly in the LLM's input embedding space, optimized with a cosine-distance loss, rather than discrete tokenization or diffusion. This is implemented in a decoupled Mixture-of-Transformers (MoT) architecture that bifurcates the backbone at an intermediate layer lsplit (≈L/2): a shared lower trunk feeds both a text-only understanding branch (cross-entropy loss) and a parallel generation branch (NEP loss via a dedicated vision head), so generation gradients enrich the shared trunk without directly touching upper understanding layers; the generation branch is discarded at inference. Training uses ~10M generation samples across five task categories (grounding, segmentation, image editing, visual chain-of-thought, text-to-image) built via an automated annotation-free pipeline (RAM++ tagging, Grounding DINO detection, difficulty-controlled sampling, LLM instruction synthesis), in a two-stage recipe (MoT Align on T2I data, then Joint Training with a λ ramp from 0.015 to 1.0) on Qwen3-VL-ViT + Qwen3 2B/4B backbones.

Key results:

• At 2B scale: DynaMath improves from 46.2 to 47.9 (+1.7pp), MathVista from 54.4 to 56.4 (+2.0pp), CountBenchQA from 87.7 to 90.1, CV-Bench-2D from 69.9 to 73.2, VisuLogic from 26.8 to 28.6.
• At 4B scale, GAS achieves MMMU 57.0, CharXiv-DS 78.4/RS 40.5, CountBenchQA 90.8, surpassing larger unified models like BAGEL (7B+7B) and Emu3 (8B) on most of 16 benchmarks.
• Per-task ablation: Segmentation and Grounding boost Count&Spatial by +2.00pp and +2.37pp respectively (CV-Bench-3D up to +9.8pp from segmentation); combining all five tasks reaches Count&Spatial 75.72, exceeding any single task, and MathVision 20.1 versus 18.1 best single-task.
• Training cost overhead is only +11.6% GPU-hours (2,464 vs. 2,208 baseline GPU-hours) with strictly zero inference overhead since the generation branch is removed after training.

Why it matters / caveats: The paper demonstrates generation-guided training as a practical, stage-agnostic (effective from-scratch, after pretraining, and after SFT) route to improving multimodal understanding at no deployment cost; however, gains are task-dependent (e.g., VisuLogic regresses at 4B scale) and ablations show applying NEP directly to a shared backbone (without MoT decoupling) causes broad degradation, underscoring that architecture and objective choices, not just added data, are necessary for transfer.

Problem: The paper investigates whether committees of LLM agents deliberating on shared clinical benchmarks can be gamed by shortcuts and socially plausible cues that a benchmark rewards but a clinician would ignore, and whether such gaming can be detected.

Method: Using Gemini 2.5 Flash and Flash-Lite agents at temperature 0 across seven cohorts on six public clinical datasets (MedQA-USMLE, MedMCQA, MIMIC-CXR reports, NIH ChestX-ray14, MIMIC-CXR-JPG, CheXpert, SUPPORT2), the authors run controlled cue-sensitivity experiments (comparing model responses on original vs. cue-modified inputs, with a repeated-query noise floor) and peer-endorsement experiments (a paired design where a holdout agent is exposed to one or two peers asserting a predefined wrong answer, or a false automated "pre-screen" flag, measuring adoption/contagion via McNemar tests). They evaluate three oversight mechanisms — a "gate" (flags every shared-condition agreement), a same-lineage "judge" (reads only the transcript), and a "referee" (privately re-queries the holdout without the transcript) — scored by precision, recall, and false-positive rate, plus a hidden-rubric experiment testing whether a decoy answer rewarded by an unstated rubric induces silent drift.

Key results:

• Individual (solo) models largely resist cues, flipping only 5–16% of cases, but two endorsing peers asserting the same wrong answer produce adoption in 38% of MedQA cases (and similarly for a false "pre-screen" system flag), with imaging cascades even stronger (e.g., NIH ChestX-ray14 contagion of +0.54 to +0.57).
• The gate has a false-positive rate of 100% (cannot distinguish adoption from honest agreement); the transcript-only judge achieves precision 1.0/recall 0.93 on text but collapses onto the gate in imaging; the referee (private re-query) achieves 77–88% precision with 13–21% false-positive rate on imaging and transfers across modalities.
• Tripling visual cue salience (opacity 0.15→0.45) does not increase contagion, while adding a second peer voice raises adoption by roughly half again; a confident peer assertion is adopted in 0.42 of cases vs. 0.14 when hedged.
• Under a hidden rubric rewarding a wrong "decoy" answer, drift is substantial (e.g., +0.83 on NIH X-rays, +0.86 on CheXpert) but almost never self-reported: only 1/10 text and 1/134 imaging "drifters" name the rubric they moved toward.

Why it matters / caveats: The authors argue for "mandatory pluralism" in clinical AI oversight — a single overseer sharing the committee's blind spot (as with the gate and judge) fails, so a referee that independently re-queries the agent is needed, and self-reported justifications cannot be trusted to catch reward-hacking. The authors also flag construct-validity issues in their own design (e.g., some referee flags reduce algebraically to their scoring label, five metric arms were withdrawn) and note the imaging cue rendering is only reproducible under a pinned library version due to a font-rasterization determinism bug discovered during replication.