AI papers — 2026-07-02

Problem: Current multimodal perception benchmarks rely on saturated, linearly-averaged, and often sparse/biased reference metrics, so they fail to reliably reflect genuine perceptual capability and real-world brittleness of multimodal large language models (MLLMs).

Method: The authors construct PerceptionRubrics, a benchmark of 1,038 information-dense images paired with over 10,000 instance-specific rubrics, built via a caption-centric pipeline: golden captions are first produced through a "Circular Peer-Review" consensus process among three top-tier MLLMs (GPT-5.2, Gemini-3-Pro, Seed-1.8), which generate, rank, and rewrite candidate descriptions over limited iterations (N≤2), followed by discard-on-divergence filtering and lightweight human verification. Gemini-3-Pro is then used as a rubric proposer to distill each golden caption into "Must-Right" (essential fact) and "Easy-Wrong" (common hallucination/fine-grained detail) rubrics using domain-specific adaptive prompts across 7 task domains (Natural Scenes, Document & OCR, Digital UI & UX, Structured Data, STEM & Expert, Logic & Puzzle, Creative & Cultural). Evaluation uses GPT-OSS-120B as an LLM judge with a "Gated Scoring" mechanism: any failed Must-Right rubric zeroes the score (gate G), while passing responses are scored by the mean accuracy over Easy-Wrong rubrics.

Key results:

• Benchmark contains 1,038 images, 10,718 total rubrics (4,053 Must-Right, 6,665 Easy-Wrong), averaging 10.33 rubrics/image; average golden caption length 770.42 words (median 569).
• 25 MLLMs evaluated; Seed-2.0-Lite tops the leaderboard at 70.07% overall, beating runner-up Gemini-3.5-Flash (69.88%) by 0.19%.
• GPT-4o-2024-05-13 scores only 12.59% overall, the weakest among proprietary models.
• Best open-source model (Qwen3.5-397B-A17B, 61.61%) trails the best proprietary model by over 8 percentage points.
• Models perform best on Natural Scenes (e.g., 79.20% for Seed-2.0-Lite) and worst on GUI domain (e.g., Qwen2.5-VL-7B drops to 5.13%).
• Near-perfect correlation (R² ≈ 0.98) between Must-Right Pass Rate and Easy-Wrong accuracy across models.
• PerceptionRubrics achieves Pearson correlation 0.916 and Spearman rank correlation 1.000 with Vision Arena human-preference Elo scores, outperforming DOCCI and DetailCaps in human alignment.
• Length-bias check: Gemini-3.1-Pro shows no significant correlation between response length and score (r = -0.079, p = 0.0758); Kimi-K2.6 shows weak positive correlation (r = 0.172, p = 1.09×10⁻⁴).
• Judge robustness: GPT-OSS-120B judge scores are systematically ~6.0% lower than GPT-5.5 judge scores but produce identical model rankings.
• Rubric coverage vs. stability: standard deviation of model scores decreases monotonically as sampled rubric ratio increases from 20% to 80%.

Why it matters / caveats: The gated scoring reveals a "Reliability Gap" where models pass most individual atomic checks but fail strict conjunctive Must-Right constraints, showing that high partial/holistic scores mask real perceptual brittleness (especially in GUI and other dense domains); this makes strict perceptual fidelity, rather than coarse semantic similarity, a better prerequisite signal for reliable multimodal generation and a more human-aligned evaluation standard. A caveat is that the golden-caption/rubric pipeline itself relies on other frontier MLLMs (as jury, generator, and rubric proposer) and an LLM judge (GPT-OSS-120B) for scoring, so benchmark quality is bounded by these models' own biases and capabilities, and samples with model disagreement during consensus filtering are discarded rather than resolved.

Problem: Existing video-generation serving systems are built for stateless, one-shot generation requests and fail to efficiently serve streaming video generation, which requires long-lived, stateful multi-user sessions with per-chunk latency targets under heterogeneous session durations and fluctuating demand across multiple GPUs.

Method: TurboServe formulates streaming video generation serving as an online, event-driven scheduling problem that jointly optimizes session placement and GPU provisioning, implemented as a closed-loop algorithm combining a migration-aware placement controller (event-driven min-max rebalancing with a gain function trading off latency reduction against an α-β-model migration cost) and a load-driven autoscaling controller (hysteresis-based scale triggers with proportional target-utilization tracking and adaptive control parameters λ(t)/ρ̂(t), inspired by the Quasar workload-classification paradigm). It is deployed atop a TurboServeBase runtime that performs coalesced chunk processing (batching concurrent sessions per GPU), GPU-CPU state offloading for suspend/resume, and NCCL/RDMA/NIXL-based GPU-GPU state migration at chunk boundaries. Evaluation uses LongLive-style autoregressive streaming video models (LongLive-1.3B and a 7B variant) served over real production traces (Trace 1-6) from Shengshu Technology on clusters of 16 NVIDIA H20 GPUs and 64 NVIDIA B300 GPUs.

Key results:

• Reduces worst-case per-chunk latency by 37.5% on average (up to 51.6%) versus baseline serving configurations under matched GPU cost.
• Reduces total GPU operating cost by 37.2% on average (up to 49.0%) versus baselines under matched latency constraints.
• On the 1.3B model, Trace 1, Cluster 2: latency reduced 28.2% vs TurboServeBase, 20.5% vs +LAG, 26.6% vs +MAG; cost reduced 38.3% vs TurboServeBase, 35.9% vs +LAG, 16.7% vs +MAG.
• Ablations: disabling migration increases GPU cost by 15.0% on average (up to 28.0%); disabling autoscaling increases GPU cost by 42.9% on average (up to 80.4%).
• Migration-aware rebalancing scheduling completes within 15 ms on clusters up to 64 GPUs (<2% of per-chunk generation time) and within 0.1 s at 256 GPUs; it closes the gap to an exhaustive-search oracle to 3.6% average (6.5% max) while cutting scheduling time by >10x.
• Autoscaling stays within 6.1% of an offline lower-bound oracle cost on average (8.3% maximum) across three traces on Cluster 2 (e.g., Trace 1: $196.87 vs oracle $188.03).
• Session migration overhead is 23-30 ms per migration, i.e., only 2-3% of per-chunk generation latency (917-1201 ms) across H20/B300 and 1.3B/7B configurations.

Why it matters / caveats: TurboServe is presented as the first serving system designed specifically for streaming video generation, showing that jointly coordinating session placement and elastic GPU provisioning is needed to control both latency and cost under bursty, long-lived session workloads; results are validated only on Shengshu Technology's internal production traces and LongLive-style models on H20/B300 clusters, so generalization to other streaming video architectures or public workloads is not directly demonstrated in the text.

Problem: Long-term memory in LLM-based agents can induce sycophancy, causing agents to over-rely on outdated or inapplicable retrieved memories instead of objective evidence, and existing memory benchmarks fail to evaluate this post-retrieval reasoning failure.

Method: The authors introduce MemSyco-Bench, a benchmark built via a four-step pipeline (memory-decision schema construction, question instantiation, long-term dialogue simulation, and multi-stage quality validation) that embeds semantically related historical memory fragments into natural multi-turn dialogues across five task categories: Objective Fact Judgment, Contextual Scope Control, Memory-Evidence Conflict, Valid Memory Selection, and Personalized Memory Use. They evaluate seven memory systems (NaiveRAG, Mem0, A-Mem, LightMem, MemGPT, MemoryBank, SuperMemory) on two backbone models (Qwen3-8B and DeepSeek-V4-Flash), reporting Generation Accuracy plus task-specific metrics (Sycophancy Rate, Correct Memory Use, Outdated Memory use), and also test lightweight interventions (a memory-caution instruction and a confirmation instruction).

Key results:

• In a preliminary study, adding incorrect memory snippets dropped DeepSeek-V4-Flash accuracy from 56.1% to 40.2% and raised its sycophancy rate from 24.3% to 52.3%.
• Across four existing memory benchmarks (LongMemEval, LoCoMo, STALE, PersonaMem), retrieval-failure errors accounted for 47.4%-66.1% of samples versus only 5.8%-13.7% for post-retrieval misuse, showing existing benchmarks mostly test retrieval, not sycophancy.
• On MemSyco-Bench, in Objective Fact Judgment, Qwen3-8B accuracy dropped from 49.12 (No Memory) to 26.00-36.00 with memory systems; DeepSeek-V4-Flash dropped from 74.33 to 56.33-63.37.
• In Contextual Scope Control, Mem0 and LightMem reduced Qwen3-8B accuracy from 70.00 to 13.34/13.67, and DeepSeek-V4-Flash from 79.00 to 28.00/33.33.
• In Memory-Evidence Conflict, Full Dialog on Qwen3-8B reached only 0.67% accuracy with a 99.33% sycophancy rate.
• Error attribution: across Mem0, A-Mem, and LightMem, 61-62% of all errors occurred after relevant memory was already retrieved (post-retrieval misuse), reaching 64%, 74%, and 75% for A-Mem on three task categories.
• A memory-caution instruction improved Memory-Evidence Conflict accuracy (Full Dialog +31.6%, A-Mem +9.8%) but hurt Personalized Memory Use by 13.0-21.0%.
• A confirmation ("Are you sure?") instruction degraded average performance by 9.9-27.7 points across settings and dropped Personalized Memory Use by 22.0-46.3%.
• On Valid Memory Selection, A-Mem retrieved both old and updated memories in 98.57% of cases but still only reached 24.06% accuracy; Mem0 accuracy fell from 53.06% (updated memory only) to 26.38% when old and updated memories co-occurred.

Why it matters / caveats: The findings show that current agent memory systems (Mem0, A-Mem, LightMem, MemGPT, etc.) frequently increase sycophancy rather than mitigate it and fail at post-retrieval arbitration between conflicting/outdated memories and objective evidence, indicating that improving retrieval alone is insufficient and that memory systems need explicit temporal/evidentiary arbitration mechanisms; the paper itself is marked "Work in Progress."

Problem: In prefill-decode (PD) disaggregated LLM serving, existing decode routers balance only per-worker load and treat decode workers as interchangeable, which is incomplete for mixture-of-experts (MoE) models because equally loaded decode workers can differ substantially in latency depending on which distinct experts their batched requests activate.

Method: ELDR (Expert-Locality-aware Decode Routing) builds a per-request "expert signature" from prefill-time expert activations (IDF-reweighted, discrete top-k expert counts restricted to a greedy-selected subset of layers, L2-normalized) that predicts decode-time expert usage; offline, Hungarian-balanced K-means clusters these signatures into one locality region per decode worker; online, "locality-band routing" routes each request to the least-loaded worker among those whose centroids fall within a similarity band (τ=0.1) of the request's best-matching centroid. A signature cache co-indexed with the vLLM KV cache at KV-block granularity keeps signatures exact under partial/full prefix-cache hits and evictions. It is implemented as a ~2,000-line addition to vLLM 0.21.0rc1 (NIXL connector) and evaluated on Qwen3-30B-A3B, GPT-OSS-120B, and Gemma-4-26B-A4B (plus a 235B expert-parallel scaling test).

Key results:

• Reduces median TPOT by 7.0–13.9% (task workload) and 5.9–10.0% (language workload) over the best of four load-balancing baselines (Random, RR, JSQ, P2C), across three MoE models at 8P16D on 24 GPUs.
• Reduces tail (P99) TPOT by 3.4–6.0% on the task workload.
• Beats a "Domain" oracle-label baseline by 1.4–6.9% (median) and 1.6–4.5% (tail) on task workloads, and by 5.7–9.1% (median) and 7.0–9.5% (tail) on language workloads.
• Cuts per-step active-expert count by 22.0% on average vs. round-robin (Qwen3-30B-A3B, task workload).
• On Qwen3-30B-A3B, growing active experts from 16 to 128 raises MoE-layer latency 4.7x at fixed batch size.
• Runtime overhead: 0.86 ms per request (1.2% of 69 ms median TTFT); signature cache uses 0.24% of HBM and under 1% of KV cache size.
• On Qwen3-235B-A22B at 2P8D (40 GPUs, 5 nodes), ELDR reduces median TPOT by 2.7–4.3% and tail TPOT by 0.6–2.0%.
• Model outputs are unchanged (lossless): ELDR only changes which decode worker serves a request, not any token's expert selection.

Why it matters / caveats: ELDR identifies expert locality as a previously unexploited, predictable routing axis for PD-disaggregated MoE serving and demonstrates a lossless, low-overhead way to exploit it that composes with existing prefix-caching and expert-parallel load-balancing systems. A caveat noted in the paper is that on the language workload the mean tail-TPOT benefit is smaller and even regresses slightly for two of three models (1.5% on GPT-OSS-120B, 0.2% on Gemma-4-26B-A4B) when averaged across rates, though at per-cell peak rates all three models still improve.

Problem: Vision-Language-Action (VLA) models trained in a source environment suffer substantial performance degradation when deployed under environmental shifts (e.g., camera viewpoint changes or embodiment changes like Panda to UR5e), and existing adaptation methods require costly multi-demonstration, per-task fine-tuning in the target domain.

Method: The authors propose Domain ARiThmetic (DART), which fine-tunes a base VLA policy (π0.5 or π0-FAST) on one-shot demonstrations of the same task in both the source and target domains to obtain "update-vectors" (parameter deltas from the base model), then extracts a domain vector by subtracting the source update-vector from the target update-vector via weight arithmetic. To purify this domain vector, DART applies SVD-based "subspace filtering" (removing source update-vector basis components misaligned with the target subspace, using a dynamic cutoff derived from a subspace alignment score γ) and "subspace scaling" (down-weighting the domain vector by γ when source/target subspaces are poorly aligned). The refined domain vector is added back to the base policy weights with a scalar coefficient α (set to 0.8) to produce the adapted policy, without any additional architecture changes or full-model retraining.

Key results:

• On LIBERO viewpoint shifts (π0.5), DART achieves average success rates of 92.0/80.8/64.4% on Small/Medium/Large shifts (79.1% average), versus 74.3% for FLA, 69.6% for RETAIN, 31.5% for One-shot FT, and 54.5% for Zero-shot.
• Under combined visual perturbations on LIBERO with π0.5, DART averages 75.0% vs. FLA 71.5%, RETAIN 68.7%, One-shot FT 29.8%, Zero-shot 60.5%.
• On π0-FAST (autoregressive architecture), DART averages 79.4% vs. FLA 76.6%, RETAIN 76.5%, One-shot FT 62.1%, Zero-shot 73.4%.
• Cross-embodiment transfer (Panda to UR5e) on MimicGen: DART achieves 84.3% average progress rate / 69.4% average success rate vs. 79.8%/62.0% (Zero-shot) and 74.4%/56.4% (One-shot FT).
• Real-world UR10e experiments (5 tasks, one demonstration for adaptation): DART achieves 81.7% average success rate vs. 55.0% (FLA), 51.7% (One-shot FT), 48.3% (RETAIN), 43.3% (Zero-shot).
• Ablation: adding subspace filtering and scaling improves average success from 78.1% (no components) to 79.1% (both).
• Merging three domain vectors (from three viewpoints) into one consolidated vector still yields 79.1% average success with DART's merging vs. 70.9-75.7% for other model-merging baselines (TA, TIES, TSV, Iso-C).

Why it matters / caveats: DART enables data-efficient (single-demonstration) VLA adaptation across visual and embodiment shifts without architectural changes, reducing the deployment bottleneck of collecting extensive target-domain demonstrations; however, the authors note performance still degrades under severe shifts (e.g., Large viewpoint change), and the scalar coefficient α requires a small hyperparameter search, with fully hyperparameter-free per-layer adaptive scaling left to future work.

Problem: Multimodal Large Language Models forced to reason through discrete text tokens suffer a language-space bottleneck that loses perceptual nuance, and existing continuous latent-reasoning alternatives introduce a train-inference mismatch because an answer-conditioned training posterior can exploit "answer leakage" shortcuts unavailable to the target-agnostic inference-time prior.

Method: The paper proposes Asymmetric Mutual Variational Learning (AMVL), which inserts k continuous latent placeholder tokens (<latent>) into the autoregressive sequence between prompt x and answer y, parameterizing a target-agnostic prior pθ(Z|x) and a target-aware posterior qϕ(Z|x,y) as factorized diagonal Gaussians via lightweight LLM-native variational heads (RMSNorm + SwiGLU projection) built on Qwen2.5-VL-7B-Instruct. Training combines the standard next-token-prediction loss with a forward KL term (trains the prior to match posterior latents) and a reverse KL term (regularizes the posterior against drifting into inference-incompatible regions), using asymmetric scheduling of the two KL weights; latents are sampled via reparameterization and injected at inference from the prior alone. The model is trained on a mixture of Visual-CoT, ReFocus, CogCoM, and Zebra-CoT datasets with k=8 latent tokens, dimension d=512.

Key results:

• Fine-grained perception (V, HRBench4K, HRBench8K average): AMVL reaches 74.97 average vs. 69.40 for base Qwen2.5-VL-7B (+5.57), with +7.85 on V (84.29 vs 76.44).
• BLINK benchmark: average score improves by +10.83 (66.91 vs. 56.08 baseline), including a +32.00 gain on Jigsaw (77.33 vs 45.33).
• Outperforms continuous latent-reasoning baselines LVR, Mull-Tokens, and Monet, and discrete reasoning baselines Vision-R1, PAPO, PixelReasoner, and DeepEyes, across virtually all sub-metrics.
• Ablation: full dual-KL objective achieves best scores (V*=84.29, HRBench4K=72.12, HRBench8K=68.50) versus NTP-only (81.15/70.50/67.38) or forward-KL-only, which collapses to 40.84/53.37/52.00.
• Latent configuration ablation: k=8 and d=512 are optimal; increasing k to 16 or d to 768 degrades performance (e.g., V* drops to 81.15 at k=16, and to 79.58 at d=768).

Why it matters / caveats: The results support the claim that resolving the prior-posterior train-inference mismatch via bidirectional (forward + reverse) KL calibration, rather than relying on hand-crafted latent supervision as in prior continuous-reasoning methods, yields a more inference-compatible and expressive latent reasoning space; the paper also provides theoretical propositions proving the dual-KL objective reduces prior contamination from answer leakage compared to standard one-sided ELBO training.

Problem: Existing automated LLM data-mixing methods (e.g., RegMix) assume a static global mapping from mixture weights to validation loss, so when the underlying data pool shifts they require costly retraining from scratch and fail to account for state-dependent optimal mixtures, especially during SFT.

Method: CausalMix casts SFT data mixture optimization as a state-conditioned causal inference problem, treating the log-transformed domain mixture as a continuous treatment, data-pool statistics (Normalized_Loss, Writing_Style, HES) as covariates X, and downstream task performance as outcome Y, using a partially linear model Y = θ(X)⊤log(T) + g(X) + ε. It applies Double Machine Learning (DML) with LightGBM as the first-stage nuisance predictor and CausalForestDML (via the EconML package) as the CATE estimator, then converts estimated marginal returns into a feasible mixture either analytically (CausalMix-A) or via search (CausalMix-S, averaging top-100 of 100,000 Dirichlet-sampled candidates). The causal model is fit on 512 proxy runs of Qwen2.5-0.5B on tulu-3-sft-mixture subsets, then extrapolated to train an 800K-data Qwen2.5-7B model and a Qwen3-4B model on the unseen AM-Thinking-v1-Distilled-Code&Math LongCoT dataset.

Key results:

• On Qwen2.5-0.5B at 800K scale, CausalMix-A reaches AvgDev 33.94 vs. best baseline DMO's 32.04 and Equal-mixture's 31.78 (RegMix 26.40).
• Scaled to Qwen2.5-7B on 800K data, CausalMix-S achieves the best AvgDev of 62.28 versus DMO 60.35, Equal 60.02, RegMix 60.14.
• On unseen LongCoT data (Qwen3-4B), CausalMix attains an overall average of 66.66, outperforming Grid (64.74), Equal (63.80), DMO (63.47), DoReMi (62.00), RegMix (61.40), ODM (58.77).
• Ablations on Qwen2.5-0.5B/800K show removing covariates drops Avg to 33.29 and removing DML orthogonalization drops Avg to 32.66, versus 33.94 for full CausalMix-A.
• Causal estimator selection: CausalForestDML scores best at R-Score +0.1683 vs. LinearDML +0.1445, SparseLinearDML -1.7065.
• Covariate selection via Spearman correlation: the 3-covariate combination [HES, Normalized_Loss, Writing_Style] achieves the best correlation of 0.7557, versus single covariate HES alone (0.7307).

Why it matters / caveats: By modeling data mixing as causal marginal-return estimation rather than a static global mapping, CausalMix generalizes to unseen data pools and larger/different model architectures without retraining proxy models, and its CATE tree interpreter reveals interpretable phenomena such as "skill conflicts" between knowledge injection and reasoning; however, the approach is validated only on a fixed set of 512 proxy runs with a small (Qwen2.5-0.5B) proxy model, and the paper notes covariate selection is limited by this modest meta-dataset size.

Problem: Vision-language models struggle with fine-grained visual reasoning because small but critical visual cues get buried in high-resolution images, and existing methods don't explicitly separate perception (locating evidence) from reasoning (answering).

Method: The paper proposes Perceive-to-Reason (P2R), a two-stage inference framework where a single shared-parameter VLM first acts as a Perceiver to predict bounding boxes for question-relevant regions, then acts as a Reasoner that answers using an annotated (box-overlaid) image plus cropped evidence patches. To train this decoupled formulation, the authors introduce Perception-Reasoning Alternating GRPO (PRA-GRPO), a role-aware reinforcement learning strategy built on GRPO that alternates between perception-focused and reasoning-focused policy updates (freezing the other role each phase), using only a binary final-answer-correctness reward without ground-truth bounding box supervision. P2R is instantiated on Qwen3-VL-Instruct-2B/4B/8B backbones, trained on a 10K-example set (from DeepEyes, VisualProbe, ZwZ) using GRPO with group size 8.

Key results:

• P2R-4B achieves 93.2% on V-Star, 81.9% on HR-Bench-4K, and 80.5% on HR-Bench-8K, substantially outperforming Qwen3-VL-Instruct-4B (81.7%, 73.8%, 67.0% respectively).
• Averaged over V-Star, HR-Bench-4K, HR-Bench-8K: P2R-2B/4B/8B improve over their backbones by +8.1%, +11.0%, +9.7% respectively; P2R-8B reaches 85.9% average, best among open-source methods compared.
• On MME-RealWorld-Lite, P2R-2B/4B/8B improve overall scores by +4.0%, +7.1%, +7.0% over backbones.
• Diagnostic study: giving Qwen3-VL-4B oracle bounding boxes raises V-Star accuracy from 81.7% to 90.6%, showing perception (not reasoning) is the main bottleneck.
• Ablation: switching from direct CoT to P2R inference raises V-Star score from 81.7% to 89.0% for the base model and from 84.8% to 93.2% for P2R-4B; alternating training order Perceiver-then-Reasoner reaches 93.2% vs. 90.6% for Reasoner-then-Perceiver.
• On ReasonSeg grounding generalization, P2R-4B improves Acc@0.5 by an average of +1.1%, despite no grounding-specific training data.

Why it matters / caveats: The results support that explicitly decoupling "where to look" from "how to reason," combined with role-aware alternating RL, is an effective and annotation-light way to improve fine-grained visual reasoning across model scales and generalizes to broader multimodal and grounding tasks. Limitations noted by the authors: the two-stage pipeline adds inference cost versus direct prompting, scaling behavior at larger training scales was not explored due to compute constraints, and the final-answer-only reward is a sparse training signal.

Problem: Existing Vision-Language-Action policies and World Action Models are poorly aligned with the structure of mobile manipulation—operating on coarse video chunks, entangling navigation and manipulation actions, and training inverse dynamics under conditions that mismatch autoregressive inference—causing loss of fine-grained contact dynamics, action-space conflicts, and compounding errors over long-horizon rollouts.

Method: ABot-M0.5 is a video-action World Action Model built on the Wan2.2 (5B) video diffusion backbone that factorizes generation into a cascade of future video latent → frame-level latent action → executable robot action, trained end-to-end with Conditional Flow Matching (CFM). It introduces intermediate latent actions (extracted via a frozen encoder pretrained with the ALAM framework) as an embodiment-agnostic bridging representation, and a dual-level Mixture-of-Transformers (D-MoT) that disentangles modality streams and further splits the action stream into mobility and manipulation sub-towers. Training follows a progressive paradigm: large-scale world-model pretraining on OXE, Agibot-Beta, RoboCOIN, RoboMind, Galaxea, and others; self-supervised latent-action pretraining; then two SFT stages, the second applying "Dream Forcing," which conditions inverse-dynamics training on the model's own self-dreamed (predicted) outputs rather than teacher forcing.

Key results:

• RoboCasa365 (pretraining protocol): ABot-M0.5 achieves 40.4% average success vs. Qwen-RobotManip 35.9%, RLDX-1 33.2%, GR00T-N1.5 23.9%, π0.5 16.9%, π0 14.8%, Diffusion Policy 6.1%.
• RoboCasa365 (Target 100% protocol): ABot-M0.5 reaches 54.2% average vs. Lingbot-VA 45.1%, GR00T-N1.5 43.7%, Fast-WAM 43.5%.
• RoboTwin 2.0 (bimanual, 50 tasks): ABot-M0.5 averages 94.10%, best among compared baselines (e.g., Qwen-RobotManip 93.85%, π0.5 79.75%).
• LIBERO: ABot-M0.5 achieves 99.4% average, edging out CORAL (99.3%) and PriorVLA (99.1%).
• Ablations: 3-stage separate latent-action design gives 94.00% success on RoboTwin-Clean vs. 87.60% baseline (direct video-to-action).
• Dream Forcing: from a 50k-step warm-start checkpoint (67.55% success), +5k steps of Dream Forcing reaches 70.56% (+3.01 points), while continued teacher forcing only reaches 66.78-68.90%.
• Pretraining effect: on RoboCasa365 Target 10%, pretrained ABot-M0.5 reaches 49.0% success vs. only 17.8% for a Wan2.2-initialized model fine-tuned directly.
• Real-world: evaluated on an Agilex Piper 6-DoF arm with only 50 demonstrations per task, outperforming π0.5 and Fast-WAM.

Why it matters / caveats: The results indicate gains stem from jointly aligning temporal granularity, action-space structure, and train-test conditioning rather than merely scaling model or data size. Caveats: RoboCasa365 composite-unseen success remains low (2.7%) even for the proposed method, indicating limited generalization to unseen task compositions, and the "condensed memory" extension is only briefly mentioned with details deferred to future work.

Problem: LLM agents currently solve competition-level or academic-style problems well but often fail to reliably complete practical, long-horizon, real-world tasks end-to-end, and existing evaluations don't adequately capture domain-specific, long-tail, agentic workloads seen in production.

Method: The ByteDance Seed team presents Seed2.0, a series of three general-purpose LLMs (Pro/Lite/Mini) built for large-scale production deployment, with a video tool-use mechanism called VideoCut that replays relevant video segments at higher FPS for long-video reasoning. The team constructs an evaluation framework spanning Fundamental Language, Vision, and Agentic capacity plus a four-dimension "Advanced Economically & Scientifically Valuable Tasks" suite, introducing new benchmarks such as LPFQA, Encyclo-K, HLE-Verified, NL2Repo-Bench, AInstein Bench, XpertBench, and GDPVal-Verified. The paper does not disclose model architecture, parameter counts, or pretraining/fine-tuning recipe details.

Key results:

• Seed2.0 Pro achieves gold-medal performance on 2025 IMO (35/42) and 2025 CMO (114/126); Codeforces Elo of 3020; Putnam-200 Pass@8 of 35.5 (vs. Gemini-3-Pro 30.5).
• On fundamental language benchmarks: HealthBench 57.7, Encyclo-K 65.7 (best among compared models), IMOAnswerBench 89.3, LiveCodeBench(v6) 87.8.
• On vision benchmarks: MathVista 89.8, MathVision 88.8, MathKangaroo 90.5 (all SOTA/tied SOTA); MMLongBench-Doc 61.4 (SOTA).
• On video benchmarks: VideoMME 89.5, Morse-500 37.4% (new SOTA), CrossVid 60.3 (SOTA); with VideoCut tool-use, ZeroVideo score rises from 14.5 to 27.9.
• On agentic benchmarks: BrowseComp-zh 82.4, HLE-Verified 73.6, τ²-Bench (retail) 90.4, SpreadsheetBench Verified 79.1 (best); but trails on SWE-Evo (8.5 vs. Claude-Opus-4.5's 27.1) and NL2Repo-Bench (27.9 vs GPT-5.2's 49.3).
• API pricing is roughly an order of magnitude lower than frontier competitors (e.g., Seed2.0 Pro: $0.47/$2.37 per 1M input/output tokens vs. Claude-Opus-4.5-thinking's $5.00/$25.00).

Why it matters / caveats: Seed2.0 aims to close the gap with international frontier models (GPT-5.2, Claude, Gemini) at substantially lower cost for large-scale deployment, but the authors explicitly acknowledge remaining gaps versus Claude on coding, versus Gemini on long-tail knowledge, and versus human performance on video motion/perception and repository-level coding tasks.

Problem: Traditional robot programming and existing coding-agent approaches struggle because they only get coarse task-level failure feedback and never accumulate reusable experience across tasks, so debugging insight is discarded after each task rather than compounding.

Method: ASPIRE (Agentic Skill Programming through Iterative Robot Exploration) is a coordinator-actor continual learning system that writes code-as-policy robot programs in the CaP-X framework (built on MuJoCo Playground), using Claude Code with Claude Opus 4.6 as the coding agent in simulation and OpenAI Codex GPT-5.5 on a real bimanual YAM robot. It combines a closed-loop robot execution engine that logs per-primitive multimodal traces for failure localization and repair validation; a continually growing skill library that distills validated repairs into reusable in-context guidance; and an evolutionary search procedure that proposes and executes populations of candidate programs per round.

Key results:

• LIBERO-Pro (macro-avg vs. strongest baseline): +77% on Object, +41.5% on Goal, +42.5% on Spatial perturbation suites.
• Robosuite bimanual handover (2A-Hand task): improved from 20% to 92%.
• BEHAVIOR-1K navigate-and-pick-up-radio: task success improved from 56% to 88%.
• Zero-shot transfer to LIBERO-Pro Long: overall reported as 31% success vs. 4% for prior methods (CaP-Agent0/π0.5).
• Ablations on LIBERO-Pro: robot execution engine alone raises macro-average success from 14% to 62%; adding evolutionary search reaches 72%.
• Real-robot cross-embodiment transfer: Lift soda can improved 13/20 → 19/20 with skills, total tokens 61.94M → 6.58M; Open/push drawer improved 0/20 → 11/20 with skills, total tokens 334.917M → 81.67M.

Why it matters / caveats: ASPIRE shows that giving coding agents fine-grained multimodal execution traces plus a persistent, self-curated skill library lets robot programming performance compound across tasks and transfer zero-shot to harder tasks and across embodiments, reducing real-robot debugging cost by up to an order of magnitude in tokens. However, the authors note it is not yet a fully autonomous real-world lifelong learner, depends on a frozen frontier LLM, is bounded by a predefined robot API, and the debug-and-evolutionary-search loop is compute-intensive.

Problem: Standard Transformers force the same hidden-state computation stream to both predict the next token and prepare persistent state (key/value entries) for future predictions, creating a role conflict that the authors hypothesize hurts language modeling performance.

Method: The authors propose the State-Prediction Separation (SPS) Transformer, which interleaves each input token with a learned dummy <predict> token, forming two streams: a persistent "input" stream that carries state (its KV entries stay in the cache) and an ephemeral "prediction" stream (its KV entries are discarded beyond a sliding window of size w=64) that emits next-token predictions via an attention-masking scheme. They pretrain models at five scales (53M to 1.678B parameters, GPT-2-style architecture with RMSNorm, SwiGLU, RoPE) on FineWeb-Edu, comparing SPS against a STANDARD Transformer and controlled ablations (2X MEMORY, DELAYED STATE, and REVERSE SPS).

Key results:

• At 1.6B parameters, SPS matches STANDARD's validation loss using 2.6x fewer training tokens (18B vs 47B pre-decay tokens).
• SPS attains the lowest FineWeb-Edu validation NLL at every scale, with the gap over STANDARD widening from -0.042 (53M params) to -0.068 (1.678B params).
• Zero-shot accuracy (averaged over ARC-Easy, HellaSwag, PIQA, SciQ, LAMBADA) improves by 2.3-3.1 percentage points over STANDARD, e.g., 63.2% to 66.3% at the largest scale.
• SPS maintains near-parity inference efficiency versus STANDARD: peak memory ratio of 1.01 and throughput within 6-10% (0.90-0.95x), versus 2X MEMORY's +75-81% memory overhead.
• SPS beats DELAYED STATE by 0.019-0.021 in validation NLL at every scale, isolating that the separation itself (not just extra compute) drives gains.
• A 3-seed statistical test confirms SPS's advantage is significant (one-sided Welch's t-test, p<0.005) over all three baselines.

Why it matters / caveats: SPS offers substantial data efficiency at essentially no extra inference memory/latency cost, which is significant given projected exhaustion of high-quality pretraining data; however, the authors note it roughly doubles per-step training compute, was tested only on a single pretraining corpus up to 1.678B parameters, and further scaling remains untested.

Problem: AI-generated biomedical analyses of disease-associated protein signals are typically delivered as static text reports or tables, which are poorly suited for research decision-making that requires inspecting evidence, comparing mechanisms, and refining hypotheses.

Method: BioInsight is a harness-centered multi-agent system (all agents built on GPT-4o) with four agents — a Planning/Search Agent, a Reasoning Agent, a Writing Agent, and a Visualization Agent — that pass typed intermediate artifacts through an evidence layer, synthesis layer, and interface layer. Pathway ranking uses g:Profiler enrichment plus BioBERT-embedding deduplication and a weighted score; publication retrieval from PubMed/Semantic Scholar is scored via a weighted lexical/semantic/citation formula. The team also built two new benchmarks, BioInsight-1k and its 100-question expert-curated subset BioInsight-100.

Key results:

• On BioASQ Phase B Exact Answer (Batch 1), BioInsight was best or tied-best on all five metrics, beating the strongest baseline by 1.9 points in factoid MRR and 4.4 points in list F-measure.
• On BioInsight-100 (0–10 expert scoring), BioInsight achieved the highest mean score of 8.62, compared against GPT-5.5, DR-Tulu-8B, Gemma-4-31B, and Qwen3.5-9B.
• In end-to-end evaluation across five diseases against 8 systems (including Claude Sonnet 4.6, GPT-5.5+Search, Gemini Deep Research), BioInsight scored highest on evidence grounding/traceability and prioritization/ranking.
• The BioInsight-without-search ablation scored consistently lower across coverage, validity, traceability, and ranking than the full system, showing explicit search materially improves grounding.
• Case study on Alzheimer's disease correctly surfaced APOE as a cross-pathway driver and SNAP25/SYT1 as a presynaptic vesicle/chemical synapse module.

Why it matters / caveats: The results support the claim that structured, provenance-preserving intermediate artifacts (rather than fluent end-to-end generation) improve traceability and usability of biomedical evidence synthesis, but the authors caution that BioInsight is intended only for research hypothesis generation, not clinical diagnosis or treatment decisions, since retrieval can be incomplete, noisy, or subject to protein-synonym ambiguity.

Problem: World models that enable Model Predictive Control (MPC) need dynamics prediction that is both fast enough for online planning and expressive enough to represent uncertain, multimodal futures, but diffusion models' iterative inference is difficult to reconcile with the low-latency demands of latent MPC.

Method: The paper introduces Value Diffusion World Models (Valdi), a TD-MPC-style algorithm that replaces TD-MPC's deterministic MLP dynamics model with a latent diffusion dynamics model (a 6-block bidirectional encoder-only transformer using velocity parameterization), jointly trained end-to-end and online with representation learning, reward prediction, and a state-value function (rather than TD-MPC's action-value function). Training jointly optimizes a diffusion loss, TD-MPC-style reward loss, a temporal-difference value loss, and a SIGReg latent regularizer; inference uses a Cross Entropy Method (CEM) solver instead of TD-MPC's MPPI. Experiments are run on a modified CarRacing environment against a parameter-matched deterministic MLP baseline.

Key results:

• Valdi has 5,390,230 trainable parameters vs. 5,795,363 for the MLP baseline.
• With a single diffusion step at both training and inference, Valdi matches the MLP baseline's control performance within run-to-run variance, evaluated over 100 fixed tracks.
• Increasing inference-time diffusion steps (up to 8, via DDIM) does not improve control performance but substantially increases visual diversity of predicted futures (measured by LPIPS).
• At the single-step default, the planner runs at over 10 Hz and the agent acts at over 30 Hz on a single RTX 4080.
• World-model horizon H=5 corresponds to a 15-step environment horizon while only requiring 5 dynamics evaluations.

Why it matters / caveats: The results demonstrate, for the first time, that a latent diffusion dynamics model can be trained end-to-end inside a value-based online MPC loop and match a deterministic MLP baseline using just one diffusion step. However, this is explicitly a preliminary proof-of-concept limited to a single simple environment (CarRacing), performance gaps are within run-to-run variance, and increasing diffusion steps degrades rather than improves control, a tension attributed to training-inference mismatch.

Problem: Autonomous post-training of language models is highly human-intensive, and LM agents left to operate in raw CLI environments struggle to reliably plan, prepare data, train, and evaluate models for self-improvement.

Method: The paper introduces AutoTrainess, an LM agent built around AutoTrainHub, a training-specialized Agent-Computer Interface organized into four closed-loop modules — data processing, training, evaluation, and logging & planning — that externalize human training expertise as explicit workflows instead of leaving the agent an unconstrained coding task. The training module fixes LlamaFactory as the sole training backend, requiring full-parameter fine-tuning with a validation run before scaling. The system is evaluated on PostTrainBench using two agent scaffolds (Codex and OpenCode) with GPT-5.4 and DeepSeek-V4-Flash as backbone models, post-training four base models across seven benchmarks under a 10-hour, single-H20-GPU budget.

Key results:

• AutoTrainess w/ GPT-5.4 (Codex) achieves 26.94 average PostTrainBench score vs. 23.21 for CLI-only (a ~15% relative improvement).
• AutoTrainess w/ DeepSeek-V4-Flash (OpenCode) improves from 12.13 (CLI-only) to 19.58 (+7.45), demonstrating generalization across backbones.
• Ablations on Qwen3-4B (Codex): removing training module drops score to 20.2 (-12.4), removing evaluation to 24.0 (-8.6), versus full AutoTrainess 32.6.
• Removing the data interface raises train action failure rate from 7.2% to 12.7%; removing the eval interface raises eval action failure rate from 7.6% to 22.8%.
• Full interface yields 111 train-to-eval handoffs with 7 retained improvements (6.3% yield); CLI-only yields 86 handoffs with 5 retained improvements.
• Behavioral analysis: DPO-style training is one of the weakest strategies (only 1/35 improving occurrences), and annealing training improves in only 5/119 occurrences.

Why it matters / caveats: The results suggest autonomous LLM self-improvement benefits substantially from externalizing human training expertise as explicit interfaces rather than relying on raw coding ability, and the benefit generalizes across backbone models and agent harnesses. Caveats: a small performance decrease (<5%) on one of the four base models tested, and weak effectiveness of certain training strategies like DPO and annealing within the agent's exploration.

Problem: Lightweight machine learning intrusion detection models for IIoT networks are almost always evaluated only within the same network/dataset they were trained on, leaving it unverified whether their reported near-perfect accuracy transfers to unseen industrial networks.

Method: Four lightweight architectures — a depth-limited DecisionTree, a three-layer SmallMLP, a Small1DCNN, and a SmallLSTM — are trained once on Edge-IIoTset using a minimal 16-dimensional common feature schema and evaluated without retraining on two independent target datasets, Gotham 2025 and WUSTL-IIoT-2021. Explainability is assessed via SHAP and Gini feature importance; adversarial robustness is tested with the HopSkipJump black-box attack; cross-domain recovery is tested via few-shot fine-tuning on fractions of Gotham data.

Key results:

• In-domain F1 on Edge-IIoTset ranges 0.971–0.972 across all four models.
• Under natural class distribution, cross-domain F1 falls to 0.18–0.28 on Gotham and 0.09–0.13 on WUSTL-IIoT-2021 — no model retains more than 29% of its in-domain F1.
• DecisionTree's top SHAP feature appears in 40.5% of Edge-IIoTset attack traffic but only 0.42% of Gotham's (96x difference) and 0.09% of WUSTL-IIoT-2021's (435x difference).
• Adversarial robustness: DecisionTree and Small1DCNN lose 0.44–0.45 accuracy under HopSkipJump attack; SmallMLP and SmallLSTM lose 0.88, dropping to ~0.06 accuracy.
• Few-shot recovery on Gotham: DecisionTree jumps from F1=0.170 to 0.638 at 25% adaptation data; Small1DCNN shows no improvement at any fraction.
• Model footprints: 4.7–7.9 KB size, 0.11–0.40 ms inference latency, training time spans 0.20s to 85.0s, a 425x difference.

Why it matters / caveats: The findings argue that within-domain accuracy is not a valid proxy for IIoT intrusion-detection deployment readiness, and that cross-network evaluation under realistic (imbalanced) class distributions should be a standard requirement. Caveats: WUSTL-IIoT-2021 lacks TCP flag data entirely, adversarial robustness was tested on a single 100-sample run without seed repetition, and few-shot recovery was evaluated only on Gotham.

Problem: Existing autonomous scientific discovery systems either search only over restricted pairwise variable relationships or require predefined research questions/seed hypotheses, and none can reflect on their own accumulated findings to guide further open-ended, multimodal exploration.

Method: The paper introduces DiscoPER, an LLM-agent framework built on a generalized PROPOSE–EVALUATE–REFLECT loop: PROPOSE generates natural-language hypotheses with accompanying executable Python statistical-test code from tabular data plus optional images; EVALUATE runs each hypothesis's code on a training split and re-executes on a held-out validation split, accepting only claims meeting effect-size, p-value, and overfitting thresholds; REFLECT periodically analyzes accepted/rejected claims to surface gaps and compound hypotheses. Default backbone is Claude Sonnet 4.5, with ablations using Claude Opus 4.6, GPT-5.4, and DeepSeek V4 Pro; the authors also build a new benchmark, iNatDisco, from iNaturalist citizen-science observations with peer-reviewed ecological ground truth.

Key results:

• On iNatDisco-800 (9 ground-truth patterns): DiscoPER recovers 8/9 patterns with a 72.7%±3% hypothesis support rate, versus 3/9 for guided baselines, and 0–1/9 for classical causal discovery methods.
• On iNatDisco-50K (12 patterns): DiscoPER recovers 8/12 patterns with 74.2%±3% support rate, vs. 2-3/12 for guided baselines.
• Ablating REFLECT drops recall from 8/9→7/9 (iNatDisco-800) and 8/12→6/12 (iNatDisco-50K).
• On classical causal-discovery benchmarks, DiscoPER achieves edge-recovery F1 of 0.83 on SACHS (vs. 0.33–0.48 for PC/GES/NOTEARS/DAG-GNN/GOLEM) and 0.86 on ASIA.
• Backbone LLM comparison on iNatDisco-800: Claude Sonnet 4.5 gets 72.7% support/8/9 recall; Claude Opus 4.6 gets 76.5% support but only 4/9 recall.
• On a synthetic 5,000-image visual benchmark, DiscoPER recovers 3/8 patterns at 54.2% support rate vs. 0/8 for HeurekaBench and 1/8 for ExperiGen.

Why it matters / caveats: Reflective, evidence-grounded accumulation lets an LLM-driven discovery agent explore a far larger hypothesis space than classical edge-based causal discovery while still grounding claims in held-out statistical validation. However, the authors note DiscoPER is limited by the scope and bias of the observed data, its discoveries require human scrutiny, and the current vision-to-statistics pipeline lacks power to validate many visual/cross-modal hypotheses.

Problem: Standard large language models produce fluent but weakly traceable reasoning on open-ended scientific hypothesis generation tasks in materials science and mechanics, making it hard to verify that final answers are grounded in coherent, causally structured intermediate reasoning.

Method: The authors develop Graph-PRefLexOR, a family of graph-native reasoning models (1.7B, 3B, 8B) initialized from Qwen3-1.7B, Llama-3.2-3B-Instruct, and Qwen3-8B, trained with a two-stage recipe combining ORPO cold-start on teacher-distilled preference pairs followed by Group Relative Policy Optimization (GRPO) with a composite graph-quality reward. Models emit a sentinel-based reasoning trace with explicit phases (<brainstorm>, <graph>, <graph_json>, <patterns>, <synthesis>) inside a <think> block. Training data mixes fineweb-edu and a domain-specific bio-silk-mech-mix-80K corpus, evaluated on a manually curated 100-question benchmark judged by Claude Opus-4.7.

Key results:

• Graph-PRefLexOR achieves 40-65% aggregate improvement over corresponding base models on Reasoning Quality, Intellectual Depth, and Reasoning Traceability, with the largest gains in traceability.
• No-thinking baseline variants show 30-50% performance reductions, confirming gains stem from explicit structured reasoning.
• The 8B model scores ~25-30% higher than the 1.7B variant across metrics.
• Semantic diversity (inter-phase cosine distance) increases ~2-3x for reasoning traces and final answers across model sizes.
• Semantic backtracking: Qwen3-8B final answers align with its own thinking trace in only 16/100 cases, while Graph-PRefLexOR-8B final answers align with its own reasoning stages in 92/100 cases.
• Test-time graph expansion (up to ~3,700 iterations): final leap-strategy graph reaches 4,419 nodes and 37,064 edges; statistical null-model checks show relational motifs enriched at z≈100-160.

Why it matters / caveats: The results suggest graph-native GRPO training improves scientific reasoning primarily by restructuring and grounding the intermediate reasoning pathway, offering a more interpretable and traceable route to AI-assisted hypothesis generation for materials design. The paper frames conclusions based on a 100-question benchmark and LLM-judge (Claude Opus-4.7) evaluation.

Problem: Large language models frequently commit "data referencing errors" (DREs)—incorrectly citing, confusing, or omitting table values during reasoning—even when they correctly parse table structure, and these errors are not captured by final-answer accuracy alone.

Method: The authors define two DRE types (Incorrect Citation and Omitted Information) and detect them automatically via an LLM-as-a-Judge framework using Claude Sonnet-3.7 with ground truth, evaluated across models from 1.7B to 20B parameters on WTQ, TableBench, FinQA, SciTab, and ToTTo datasets. To reduce DREs they use critic-based filtering (sampling N=8 responses, selecting the subset with fewest DREs) and segment-level rejection sampling. They then train a lightweight Critic-4B model (Qwen3-4B-Instruct) via SFT on Sonnet-3.7-labeled distillation data followed by RLVR on samples derived from Qwen3-8B's WTQ responses.

Key results:

• DRE rates range widely across models on WTQ, e.g., Qwen3-1.7B 35.52%, Qwen3-8B 14.04%, gpt-oss-20b 5.71%, Llama4-Scout 46.48%.
• Prompting to avoid miscitation barely helped: Qwen3-8B DRE rate only dropped from 14.04% to 12.50% on WTQ.
• Correct-in-DRE ratio (correct final answer despite containing DREs) reached 65.57% on SciTab, showing errors persist even in "correct" responses.
• Critic-based filtering improved accuracy substantially on the DRE subset, e.g., Qwen3-8B on WTQ: 64.59% to 73.49% with CF + majority voting.
• Trained Critic-4B achieved an average F1 of 78.16% for DRE detection versus 69.51% for untrained Qwen3-4B-Instruct baseline.
• Using Critic-4B for rejection sampling still improved accuracy and reduced DRE rate, e.g., Distill-Qwen-7B on WTQ: DRE rate dropped from 46.04% to 30.59%.

Why it matters / caveats: The findings show DREs are pervasive across model sizes, families, table formats, and tasks, and are largely avoidable (simple resampling with a critic fixes them) rather than fundamental reasoning limitations. The study is limited to table-related tasks and does not investigate the underlying interpretability causes of DREs.

Problem: Existing automated metrics for medical report generation fail to reliably capture clinically critical factual errors because they either rely on surface-level n-gram overlap or are limited to narrow modalities and offer no interpretable, per-finding error attribution.

Method: AtomiMed decomposes each medical report into a two-level hierarchy of Atomic Clinical Facts (Disease-level presence/absence QA pairs and Attribute-level descriptor QA pairs) using Qwen3-235B-A22B as the decomposition and evidence-reading backbone. It then runs a bidirectional Agentic Cross-Verification loop that poses each report's questions to the other report, computing disease- and attribute-level precision/recall/F1 combined into a final F1 score. The authors also curate OmniMRG-Bench, a multi-modal benchmark (X-ray, CT, MRI, Ultrasound) with over 178K expert-verified QA pairs across 9 anatomical systems.

Key results:

• On radiologist error-count correlation, AtomiMed achieves Spearman's ρ=0.806 on ReXVal (vs. GREEN 0.798) and ρ=0.744 on ReFiSco-v0 (vs. GREEN 0.709).
• In pairwise preference study, AtomiMed achieves 95.71% ranking accuracy on X-ray versus GREEN's 63.57%.
• AtomiMed maintains 84.33% accuracy on CT and 68.19% on MRI, while GREEN's correlation collapses (τ=0.3283 on CT, 0.1513 on MRI).
• OmniMRG-Bench comprises over 178K disease- and attribute-level QA pairs, dominated by location (~62.9K) and size (~31.5K) descriptors.
• Models score much higher on Morphology (6.0–13.2) than Severity (1.3–5.9) or Size (1.0–6.9) across evaluated systems.

Why it matters / caveats: AtomiMed provides modality-universal, interpretable per-finding error attribution that generalizes beyond chest X-ray (unlike GREEN, whose correlation collapses on CT/MRI). The paper notes future work is needed to reduce inference cost via distilled backbone models and to extend coverage to longitudinal imaging comparisons and additional clinical specialties.

Problem: Existing agentic slide-generation systems handle page-level design passively (via fixed templates or lengthy user instructions) and therefore fail to capture users' latent, fine-grained design intents, leaving Page-level Slide Personalization unresolved.

Method: The paper formulates page-level slide personalization as inverse planning and proposes Spire (Structural Planning via Inverse REconstruction). Spire corrupts gold slides via random element-level structural perturbations to create self-supervised (perturbed-slide, gold-slide, discrepancy-list) triplets, then trains two Qwen2.5-VL-7B-Instruct agents — a critic and a planner — using the DAPO reinforcement learning algorithm, with reward signals from a GPT-4o-mini judge and a Claude Opus 4.5 VLM judge, and GPT-5 used as an oracle for training targets. At inference, the trained planner and critic collaborate with a black-box executor (GPT-o4-mini generating python-pptx code), trained/tested on 200 decks from Zenodo10k with SlideBench as out-of-distribution test data.

Key results:

• On test pages, Spire achieves VLM-judge average 0.5415 vs. AutoPresent (GPT-based) 0.5069, and visual-similarity average 0.7576 (vs. 0.3448 for untrained base model).
• On OOD pages (SlideBench), Spire achieves the highest judge average of 0.7333, versus AutoPresent's 0.6461.
• Ablation: removing fine-tuning drops visual-similarity average to 0.3448 and judge average to 0.3230, versus Spire's 0.7576/0.5415.
• Swapping the untrained o4-mini critic for the Spire-trained 7B critic raises visual similarity average from 0.8062 to 0.9000 and judge average from 0.4784 to 0.5035.
• PPTAgent suffers particularly poor faithfulness (0.1036 test / 0.0839 OOD) due to imperfect reference templates.

Why it matters / caveats: The results show that a small, RL-trained 7B critic/planner pair can match or beat much larger GPT-based pipelines on personalized, page-level slide design without differentiable access to the black-box rendering executor. A noted caveat is that visual-similarity metrics and judge-based quality scores are not well aligned, underscoring that pixel-level reconstruction metrics are unreliable proxies for true personalized design quality.

Problem: Multi-turn visual reasoning agents (MLLMs) repeatedly fail to localize tiny targets in high-resolution images because a single model is forced to entangle imprecise grounding with reasoning, producing long, redundant search trajectories and "inattentional blindness" (finding the right region but failing to recognize the target).

Method: PixelEyes decouples perception from reasoning by pairing a general-purpose VLM (Qwen-3-VL, 4B/8B) reasoner with an external referring-segmentation tool, SAMTok, that returns pixel-precise masks (with a fallback to bounding-box cropping via "Switchable Tool Use"). It introduces Semantic-Region Breadth-First Search, where the model anchors all coordinates to the original image and proposes new low-IoU regions whenever SAMTok fails to ground the target. Training uses a synthesized PixelEyes-6K dataset distilled from a tool-augmented Gemini-3-Flash teacher, followed by SFT then reinforcement learning with vanilla GRPO.

Key results:

• PixelEyes-8B: 94.24% on V* (vs. Qwen-3-VL-235B baseline at 87.96%), 85.00% on HR-Bench-4K, exceeding the much larger Qwen-3-VL-235B model.
• PixelEyes-4B improves over its Qwen-3-VL-4B base by +11.52 (V*), +7.00 (HR-Bench-8K).
• On the new Pinpoint-Bench (zero-hint protocol): Qwen-3-VL-4B/8B reach only 46.19%/49.88% accuracy; Mini-o3 attains LSR=78.52% but accuracy only 44.34%, showing a 34-point gap (inattentional blindness).
• PixelEyes-4B reaches 54.73% accuracy on Pinpoint-Bench.
• Ablations: fine-tuning on Mini-o3-style data drops VisualProbe-Hard to 24.52, while fine-tuning on PixelEyes-6K raises it to 50.94; adding GRPO RL further lifts it to 54.72.
• SAMTok achieves a 99.17% tool Invoke Success Rate; replacing it with Sa2VA drops Pinpoint-Bench accuracy from 54.73 to 46.19.

Why it matters / caveats: Decoupling perception (mask-precise grounding) from reasoning substantially improves both accuracy and search efficiency on needle-in-a-haystack visual search without eroding general multimodal reasoning ability. A caveat noted by the authors is that performance still leaves large headroom on Pinpoint-Bench, and grounding quality is critically dependent on the chosen segmentation backend.

Problem: Existing real-time online 3D scene graph generation methods either rely on computationally heavy SLAM-based mapping or, in the case of FROSS, approximate each object as a single 3D Gaussian, which discards fine geometric detail and causes unstable/incorrect merging of object candidates across views.

Method: NoPA builds on the FROSS pipeline (pretrained RT-DETR-EGTR 2D scene graph detector lifted into 3D via depth and camera pose) but replaces each object's single-Gaussian representation with a fixed-size non-parametric particle set, treated as samples of a kernel density estimate (KDE). Association/merging uses a two-stage rule: a Hellinger-distance pre-filter for clear cases, followed by a Maximum Mean Discrepancy (MMD) test between KDEs for ambiguous pairs; merged particle sets are resampled to keep constant memory. A relationship propagation mechanism reuses MMD scores to recover missing relation edges via majority voting.

Key results:

• On 3DSSG: NoPA (n=256) achieves 53.2% relationship recall, 69.0% object recall, 61.4% predicate recall, vs. reproduced FROSS at 25.7%/60.6%/30.7%, with comparable latency (27ms vs 22ms) and VRAM.
• NoPA with only n=128 particles (49.9% relationship recall) already surpasses all baselines including JointSSG and Kim's framework.
• On ReplicaSSG: NoPA (n=256) reaches 36.9% relationship recall vs FROSS's 22.3%, a 65.5% relative increase.
• Ablation: swapping Gaussian for particle-set representation alone actually drops relationship recall to 17.6% (from FROSS's 25.7%); adding the tailored MMD-based merging raises it to 26.3%; adding relationship propagation yields the full 53.2%.
• With ground-truth 2D input as oracle, NoPA reaches 84.4% relationship recall, versus FROSS+GT's 44.6%.

Why it matters / caveats: The non-parametric particle/KDE representation with MMD-based merging substantially improves online 3D scene graph quality over the Gaussian-based FROSS baseline while preserving real-time speed and similar memory footprint. However, performance is capped by the accuracy of the underlying pretrained 2D scene graph detector, and misclassifications still cause failures.

Problem: The paper investigates how avid readers actually experience and evaluate AI-translated ("MT") versus professionally human-translated ("HT") literary fiction, since existing automatic MT metrics and standard human evaluation protocols fail to capture immersion and literary effect.

Method: The authors built LAIT, a dataset of 15 recently published (2025–2026) French/Polish/Japanese novels translated into English, comparing published human translations against machine translations produced by an agentic LLM pipeline combining Claude Code with Claude Opus 4.6 and Codex with GPT-5.4 in a three-stage process (preprocessing/style analysis, chunk-level translation with parallel quality reviews and an acceptance gate, and excerpt-level global revision), adapted from the AutoFiction long-form fiction generation framework. Fifteen avid readers each read excerpts in both immersive (whole-excerpt) and close (300-word chunk, side-by-side) reading conditions, yielding 60 excerpt-level and 772 chunk-level evaluations plus span-level highlight annotations; automatic metrics (MetricX-QE, CometKiwi, LiTransProQA) were also run for comparison.

Key results:

• Readers preferred HT at the excerpt level in 19/30 immersive-reading comparisons (63.4% order-adjusted, p=.148, not significant) and more clearly at the chunk level (76.2% order-adjusted, p=.011, significant).
• HT had 4.0x the odds of a higher acceptability rating (p=.0069) and 4.3x the odds of a higher smoothness rating (p=.0029) than MT.
• Readers could not reliably distinguish MT from HT: 34/60 correct in single-reading (chance level, p=.413); in 28/30 comparisons readers judged their preferred version to be the human one, regardless of which was actually MT.
• HT received more positive highlights (107.8 vs. 68.5 words/1K) and fewer negative highlights (42.9 vs. 100.7 words/1K) than MT.
• Automatic metrics all favored MT over HT (e.g., LiTransProQA 0.996 vs 0.920) and correlated weakly-to-negatively with human chunk-level judgments (Kendall's τ from -0.075 to -0.318).

Why it matters / caveats: Literary MT from a strong agentic pipeline is now hard for readers to detect and often "fine," but readers still consistently favor human translation for immersion and literary quality, and standard automatic/LLM-judge metrics actively mislead by favoring MT. Caveats: evaluation covers only ~8K-word novel openings from 2025–2026 releases, each excerpt was rated by only two readers, and the multilingual case study was small and exploratory.