AI papers — 2026-08-07

Problem: In long-horizon multi-turn agentic RL with sparse trajectory-level rewards, group-relative methods like GRPO broadcast a single outcome advantage uniformly across all turns, failing to credit the few pivotal decisions that actually drive success or failure.

Method: AgentOPSD is a critic-free, turn-level credit-assignment method built on GRPO. At each turn it computes a teacher-student token log-probability gap (self-distillation contrast between a policy conditioned on a training-only retrieved "skill" and the unconditioned policy), sums these into a turn-level evidence score, and recursively updates a Bayesian belief state in log-odds space (initialized from the GRPO group success rate, with a geometric decay factor γ on accumulated evidence). The marginal belief revision per turn (signed by the outcome verifier) is normalized within-trajectory and used as a bounded multiplier (parameters b, λ) that reshapes the standard GRPO sequence-level advantage per token, plugged into a PPO-clipped objective with KL penalty — no extra rollouts or learned critic required. Trained on Qwen2.5-3B/7B-Instruct with skills retrieved from SkillRL's SkillBank (training-only; not used at inference).

Key results:

• Reaches 89.1% success rate on ALFWorld with Qwen2.5-7B, versus 81.2% for GRPO and 85.9% for the strongest self-distillation baseline (SDAR).
• Outperforms GRPO+OPSD, Skill-SD, and RLSD on all eight aggregate comparisons across two model scales (3B/7B) on ALFWorld/Search-QA/WebShop, and beats SDAR on six of eight.
• On horizon-robustness (success points lost per additional turn, Qwen2.5-7B/ALFWorld), AgentOPSD is flattest at −0.54 vs. −2.91 for GRPO and −3.59 for RLSD.
• Ablations (ALFWorld, Qwen2.5-7B, full method = 89.1): per-token instead of turn-level aggregation drops to 85.9; raw local gap instead of recursive belief revision drops to 82.8; dropping the signed direction (magnitude only) drops to 80.5; removing the empirical-rate prior anchor drops to 78.9.

Why it matters / caveats: The method restores per-turn credit assignment in the critic-free, group-relative RL setting at the cost of only one extra teacher forward pass, avoiding the additional rollouts or learned value networks required by prior long-horizon credit-assignment approaches (e.g., VinePPO, GAE). The paper notes Bk is only a "relative support" signal, not a calibrated success probability, since it is estimated via the self-teacher rather than ground truth.

Problem: Vision-language models are increasingly used as judges/reward models to verify whether computer-using agent (CUA) trajectories fulfilled their instructions, but no prior work has systematically measured whether these VLM judges are actually reliable across platforms.

Method: The authors build OSReward, a benchmark of 1,019 human-gold-labeled CUA trajectories collected on self-built cross-platform infrastructure (Web, Windows, Ubuntu, mobile) with real applications, live websites, and rich initialized states; instructions are human-written and peer-screened, and each trajectory is labeled by three independent annotators with disagreements resolved via senior meta-review. They derive OSReward-Hard (284 trajectories concentrating annotator-disputed hard cases) and OSReward-Multi (440 successful trajectories with alignment/efficiency sub-labels). They then benchmark 27 VLM judges (GPT, Claude, Gemini, Qwen, Doubao, Kimi, Intern families) under a fixed protocol, and build OS-Shepherd-100K, a 100K-sample reasoning-annotated corpus labeled via ensemble VLM-judge agreement filtering, on which they train OS-Shepherd-9B and -35B (from Qwen3.5) using a two-stage SFT-then-GRPO recipe, with the RL stage specifically targeting false-success errors.

Key results:

• Best judge (Claude-Opus-4-8) reaches 89.7% binary accuracy on the full OSReward set but drops to 69.7% on OSReward-Hard; the mean judge falls to 52% accuracy on OSReward-Hard (near a constant always-fail baseline on the 30/70 split).
• Every judge family shows a lenient bias: over-accepting an incomplete task as success accounts for at least 48% of each judge's errors, and pooled over-accepts outnumber over-rejects roughly 3:1.
• Dropping the per-step thought/action text (keeping only screenshots) costs 7.2 pp average accuracy and flips 22.7% of individual verdicts, versus <1 pp and <7% flips for various visual-input ablations — text history, not screenshots, drives verdicts.
• OS-Shepherd-9B improves over its Qwen3.5-9B base from 76.7%→86.1% accuracy (full set) and 39.4%→60.2% fail-recall on OSReward-Hard, at $1.36 per 1,000 trajectories versus ~$45–100 for frontier judges (a 30–60× cost reduction); OS-Shepherd-35B-A3B gains little over the 9B (62.7% vs 60.2% on Hard fail-recall).
• On three independent held-out benchmarks (OSWorld, WebArena, AndroidWorld), OS-Shepherd-9B beats all general open models up to Qwen3.5-397B-A17B (~44× its size).

Why it matters / caveats: The paper argues no single judge is currently both reliable and affordable, so open-sourcing OSReward and OS-Shepherd narrows that gap for large-scale CUA training/evaluation. Caveats acknowledged by the authors: quality/fine-grained grading (alignment, efficiency) remains far weaker than binary outcome judging even for the best models (low-60s% AUC vs. ~90% binary accuracy), ensembling/voting does not improve reliability because judges "herd" on the same mistakes (pairwise Cohen's κ≈0.71), and human-written verifiers themselves are imperfect, so reported judge-vs-verifier agreement likely understates true judge accuracy.

Problem: Generating large-scale, freely explorable 3D open worlds from open-ended text prompts is hard because a system must simultaneously maintain global spatial coherence, produce rich local content, and output explicit, editable/reusable 3D assets — a combination existing procedural, video-lifting, native-3D-diffusion, and MLLM-agent methods each fail to fully achieve.

Method: WorldClaw is a fully agentic, coarse-to-fine, three-stage pipeline built on Claude Opus 4.8 as the underlying agent model with task-specific agent "skills" wrapping pretrained tools (GPT-Image-2, SAM3, SAM3D, Hunyuan3D 2.1/2.5) and Blender (via BlenderMCP). Stage 1 (intent analysis and planning) converts the prompt into a structured scene specification (regions, terrain, object, and material constraints). Stage 2 (global terrain generation) uses a terrain-planning agent to produce a structured terrain spec, generates a semantic layout map plus reusable 3D asset prototypes and generative/procedural materials, builds a composite height field from region-weighted noise and geomorphic operators, scatters terrain assets, and iteratively refines geometry/materials via a render-inspect-edit loop. Stage 3 (regional object generation and placement) renders each selected region's terrain, uses an image-editing model to compose objects onto it, segments individual instances with text-guided SAM3, reconstructs each as a 3D mesh with SAM3D (with image-space scale calibration), recovers 3D placement via a two-camera ray-intersection geometric procedure, and runs an agentic refinement loop (via Blender/MCP) that fixes object pose/scale/mesh quality and resolves object-terrain contact (floating/penetration) through local co-deformation.

Key results: Not stated — the paper reports only qualitative results (four generated worlds shown via global/regional/walk views) and a qualitative comparison against SynCity, Marble, MajutsuCity, WorldGen, and GPT-5.6 Sol on terrain organization, content richness, and free-viewpoint appearance; no quantitative metrics (e.g., FID, user study scores, geometric fidelity numbers) are reported anywhere in the paper.

Why it matters / caveats: WorldClaw is positioned as producing explicit, independently editable textured meshes with explicit terrain placement (unlike Gaussian-splat or video-lifting methods, which the authors say degrade at distance or lack instance-level editability), aiming toward production-ready, game-engine-compatible world generation. However, since evaluation is purely qualitative with no held-out quantitative benchmark, the reported advantages over baselines rest on the authors' own visual comparison rather than measured metrics; the paper does not report a formal Limitations section with concrete failure numbers.

Problem: Training LLM agents for long-horizon tool use with RL typically requires either costly, hard-to-verify executable environments or external LLM simulators that are inaccurate and still need real-environment grounding, keeping environment modeling external to the acting policy.

Method: EnvACE assigns a single shared policy two roles — ACT (issue a tool call, conditioned on history) and REHEARSE (generate the environment response that the call would induce, conditioned on history and the action) — so trajectories unfold entirely within the policy without querying an external environment ("world rehearsal"). Both roles are jointly optimized end-to-end with a Role-Wise GRPO objective: each role's outputs across a group of K rollouts get a separate reward baseline (mean reward of outputs from that role), advantages are computed relative to the role-specific baseline, and the clipped PPO/GRPO objective updates the shared parameters θ from both roles' outputs. At test time, EnvACE performs N private rehearsal attempts (parallel or sequential, with self-evaluation feedback) before a single committed real-environment execution, summarizing rehearsals into a "rehearsal memory" that conditions the final action. Trained from Qwen3-8B (and 1.7B) on the CM2 dataset using verl, 470 steps, batch size 16, 4 rollouts/prompt, on 16 H20 GPUs.

Key results:

• Overall score of 32.91% (mean across BFCL V4, τ2-Bench, VitaBench averages) for EnvACE-8B, beating EnvScaler-8B by 0.99 pts and AWM-14B by 0.37 pts among environment-scaling baselines.
• On τ2-Bench, EnvACE-8B reaches 36.7% average, exceeding EnvScaler-8B (+3.8), AWM-8B (+5.5), and AWM-14B (+6.0).
• On FinMCP-Bench, EnvACE-8B achieves the best TF1 of 46.78% (vs. 43.68% for EnvScaler-8B, 42.50% for AWM-8B) and highest tool precision of 54.04%.
• Ablation vs. standard GRPO at 8B scale: EnvACE improves τ2-Bench average from 31.2% to 36.7% (+5.5 pts); a parameter-sharing ablation (separate policies for acting/rehearsal) reaches only 35.5%, showing shared parameters add +1.2 pts.
• Test-time scaling (N=2 rehearsals, parallel mode) raises Overall score from 36.7% (non-TTS) to 40.9%, with τ2-Bench average rising from 31.4% to 38.0%.

Why it matters / caveats: EnvACE offers a way to scale agent RL training and test-time deliberation without real-environment interaction, and case studies show it anticipates tool-call failures and avoids invalid write operations before execution. The authors state two explicit limitations: evaluation is only up to 8B scale due to compute constraints, and the evaluation focuses primarily on tool-interactive tasks rather than a broader range of agentic settings; they also note performance regresses slightly at N=3 rehearsals versus N=2, likely due to context-length limits.

Problem: Existing spatial-reasoning benchmarks for VLMs test local spatial perception from single or a few viewpoints, leaving it unclear whether models can consolidate long-horizon egocentric video into a globally consistent spatial representation of an entire scene.

Method: The authors build GST-Bench, a VQA benchmark with 2,762 human-verified questions (6,790 minutes of cumulative video) generated by an automatic pipeline on OmniGibson/BEHAVIOR-1K (plus HyperSim and ArtVIP for training data) across 50 indoor scenes. The pipeline produces exploration videos, object-annotated videos, off-trajectory "current view" query images, and three abstraction levels of top-down images (photo-realistic, occupancy-map, floor-plan), then generates QA pairs from ground-truth 3D geometry across 12 subtasks under three competencies: self localization (global orientation, global position), object localization (egocentric direction/distance/global position, in semantic and visual target-specification modalities, with the target guaranteed invisible from the query view), and scene structure understanding (top-down selection at 3 difficulty levels, trajectory selection). Metrics are Mean Relative Accuracy (distance), thresholded angular accuracy (orientation), thresholded pixel-distance accuracy (point prediction), and standard accuracy (multiple-choice). The authors also release GST-Train, a larger training set from the same pipeline, and fine-tune Qwen3-VL-8B on it via SFT mixed with general multimodal instruction data.

Key results:

• The best zero-shot model, Gemini-3-Pro, scores 42.68 overall vs. a human baseline of 79.08 (a 36.4-point gap); random guessing scores 20.01.
• The gap is largest on orientation estimation (21.52 vs. human 85.00) and global position for visual targets (42.23 vs. human 93.00).
• Best open-source model, InternVL3.5-38B, reaches only 30.71, with 4 of 17 open-source/embodied models (e.g., LLaVA-OV-1.5-8B at 19.58) scoring below the random baseline.
• Embodied-tuned models do not outperform general-purpose backbones of similar size: RoboBrain2.5-8B (24.61) and Cosmos-Reason2-8B (21.64) both score below general-purpose Qwen3-VL-8B (25.89).
• Controlled local-perception variants show a 2x-3x score jump for proprietary models when cross-frame reasoning is removed (e.g., Gemini-3-Pro's Egocentric Direction score rises from 22.11 to 61.20, +39.09), but open-source models show inconsistent or negative gains (e.g., InternVL3.5-8B drops by 8.03 points), indicating they fail at both local perception and global integration.
• Fine-tuning Qwen3-VL-8B on GST-Train raises its GST-Bench score from 25.89 to 53.52, surpassing all zero-shot proprietary models evaluated, though still well below the 79.08 human baseline.

Why it matters / caveats: The results show current VLMs — regardless of scale, family, or embodied-specific post-training — largely lack the ability to consolidate long-horizon video into a global spatial representation, and that this specific deficit (rather than basic local perception) is the primary bottleneck for proprietary models. The authors note the fine-tuning results demonstrate the gap is "narrowed" but not "fully solved" by targeted supervision, leaving a substantial residual gap to human performance.

Problem: Existing Reasoner–Embedder methods for unified multimodal retrieval generate Chain-of-Thought (CoT) rationales from the query alone, so they explain what the query describes rather than diagnosing what the embedder actually confuses among semantically similar candidates.

Method: UniME-R1 is an Embedder–Adviser framework built on Qwen3-VL (2B and 4B) backbones. A dual-mode embedder first retrieves top-k candidates via a discriminative embedding (<dis_emb>); a retrieval-aware adviser then analyzes candidates individually and either directly reranks them (if the target is likely already in top-k, via <rerank_judge>/<rerank_list>) or produces Retrieval-Centric CoT (<cot_focus> + <cot_answer>) that is appended to the query and re-encoded with a generative embedding (<gen_emb>) for full-corpus re-retrieval. Training uses LVLM-mined hard negatives (via an off-the-shelf embedder + false-negative filtering + LVLM-as-judge scoring) for joint contrastive training of the embedder (Ldis + Lgen), supervised fine-tuning of the adviser on teacher-annotated structured outputs, and GRPO reinforcement learning with four rewards (format, path judgment, NDCG-based rerank, and RC-CoT re-retrieval MRR-based reward).

Key results:

• On MMEB-V2, UniME-R1 reaches overall scores of 69.9 (2B) and 70.3 (4B), beating the strongest baseline in each size group by 3.1 and 1.4 points respectively, and beating Reasoner–Embedder baselines UME-R1/TTE by 9.8/6.8 points (2B) and 5.8/1.7 points (4B).
• On general multimodal retrieval (Recall@1), UniME-R1-2B improves over UniME-V2-7B by 3.1/1.1 (Flickr30K), 7.3/3.2 (COCO), 2.7/2.8 (ShareGPT4V), and 2.3/1.6 points (Urban1K); UniME-R1-4B gains 1.6 points on UVRB over Embed-RL-4B.
• Ablations: using actual retrieved candidates for RC-CoT (vs. query-only or random-candidate CoT) raises overall MMEB-V2 score from 65.6 to 68.5; adaptive rerank-or-retrieve routing (69.9) beats always-rerank (69.1) and always-retrieve (69.0), with oracle routing at 72.2; full GRPO improves over SFT-only adviser from 68.8 to 69.9.
• Inference efficiency: candidate-side latency is 0.01s vs. 0.27s for Embed-RL (27x faster), since candidates are encoded once and RC-CoT is generated only for queries.

Why it matters / caveats: The rerank-or-retrieve mechanism avoids costly candidate-side CoT generation and reuses a fixed candidate index, making retrieval-feedback reasoning practical at scale; however, the authors note routing still has a 2.3-point gap to oracle routing, and GRPO-based routing shows mild over-routing to reranking on visual-document tasks.

Problem: Multimodal LLMs struggle with multi-step temporal/visual reasoning because text-based Chain-of-Thought cannot precisely articulate continuous visual transformations, and existing multiple-choice VQA benchmarks let models exploit linguistic shortcuts instead of genuine visual reasoning.

Method: ChronoVision (a 9B MLLM) is trained in two stages. In supervised fine-tuning, a Reconstructive Visual Head (a two-layer MLP) predicts the latent visual representation of the final chronological frame from shuffled candidate-frame hidden states (trained with MSE loss added to the autoregressive cross-entropy loss), while a Region of Interest (ROI) Attention Locating module condenses text-to-image attention at a designated layer toward ground-truth bounding boxes via an Attention Condensation loss. In post-training, GRPO reinforcement learning is applied with a composite reward combining a sparse outcome-match reward, a Latent-Grounding Process reward (cosine similarity between step-wise reconstructed latents and candidate visual features), and an unsupervised visual Focus reward (negative entropy of self-attention over visual tokens). The authors also introduce Vbvr-VQA, a benchmark reformulating video reasoning as strict 6-frame chronological ordering built on the Very Big Video Reasoning dataset across 100 task generators.

Key results:

• ChronoVision achieves 74.8% in-domain and 71.6% out-of-domain accuracy on Vbvr-VQA (overall 73.2%), versus the strongest proprietary baseline Claude Opus 4.6 at 50.8% ID / 60.8% OOD, and outperforms larger open models Qwen 3.5 397B (49.4 overall) and GPT o3 (46.0 overall).
• On the out-of-domain IntPhys2 benchmark, ChronoVision reaches 55.0% overall accuracy vs. 48.5% for the Qwen3.5 9B base model (+6.5 points), with the largest gain (+11.5 points) on the Easy subset.
• Ablation on training stages: adding the Reconstructive Visual Head improves overall accuracy from 66.0% (SFT w/o head) to 69.0%; adding ROI Attention Locating raises it to 70.2%; adding GRPO RL raises it further to 73.2%.
• Ablation on reward components: removing the outcome reward drops overall accuracy by 3.4 points, removing the Latent-Grounding reward by 2.2 points, removing the ROI/Focus reward by 1.4 points.
• On 7 standard VLM benchmarks (e.g., MMMU, MathVista), ChronoVision performs comparably to vanilla Qwen 3.5 9B (e.g., 78.8 vs 78.4 on MMMU), indicating no degradation of general capability.

Why it matters / caveats: The approach shows a 9B model can outperform much larger and proprietary frontier models on temporal/visual-imagery reasoning without relying on natural-language articulation of visual transformations. The authors note the study is limited to a moderate-scale 9B setting and relies on dense auxiliary supervision (semantic locate cues and bounding-box annotations) that future work would need to reduce for scalability.

Problem: There is no shared CS/AI-systems implementation blueprint for Economic World Models (EWMs) — generative engines that simulate how economies evolve from heterogeneous agents' beliefs, actions, and institutional interactions — leaving unclear how to build, classify, and evaluate such systems as computational environments.

Method: The paper is a position/survey paper rather than an empirical model paper. It formalizes an EWM as a state-transition system (belief update, next-action prediction, agent state update, endogenous aggregation, and institutional rule evolution operators) and proposes a six-level capability ladder (L1: fixed rule-based agent worlds; L2: adaptive rule-based agents; L3: LLM-based autonomous agents; L4: self-evolving agents; L5: agent-environment co-evolution with evolving institutions; L6: sim-to-real economic twins with online correction). It then conducts a systematic literature survey: papers from arXiv (8 categories) and UTD-24-listed journals (1950–April 2026) are filtered by keyword intersection, screened in two LLM-assisted stages (GPT-5.4-mini for title/abstract, GPT-5.5 for full-text classification) and assigned to the highest supported capability level, with borderline cases manually verified. It also proposes a reusable software protocol/API (ewm.make, world.reset, world.run_agents, world.step, world.coevolve, world.align, world.evaluate) for implementing EWMs.

Key results:

• Starting from 7,836 candidate papers (6,008 arXiv + 1,828 UTD-24), stage-one LLM screening reduces the pool to 794 arXiv + 87 UTD-24 papers for full-text review, yielding a final corpus of 737 validated EWM papers.
• The literature distribution shows most work concentrated at L1–L3 (especially post-2018 on arXiv, with a sharp rise around 2024–2025), while L4–L6 systems occupy only a thin fraction of the corpus.
• UTD-24 (business/economics journal) papers are concentrated almost entirely in L1 and L2, contrasting with arXiv's push into LLM-based (L3) agents.
• L6 sim-to-real economic twins are described as "nearly absent" in the literature, with only very limited evidence of repeated real-world correction.

Why it matters / caveats: The paper argues that EWM research has advanced agent intelligence (L1–L3) but has not yet solved the harder systems problem of self-evolving agents, endogenous institutions, and real-world alignment together (L4–L6) — framing this as the key frontier for building AI training/planning/safety substrates and policy sandboxes. As a position paper, it presents no empirical validation of its proposed architecture or protocol; the capability-ladder classification also relies on LLM-assisted paper screening (GPT-5.4-mini/GPT-5.5) rather than fully manual review.

Problem: On-Policy Distillation (OPD) and its variant OPD2 (which use teacher token-level probability signals instead of RL for LLM post-training) have been studied mainly on English reasoning benchmarks, leaving their effectiveness for multilingual (Korean/Japanese) math reasoning underexplored.

Method: OPD trains a student policy πθ by sampling its own rollouts and minimizing token-level KL divergence to a teacher model π (equivalent to a token-level reward Rt = log π(yt) − log πθ(yt)). OPD2 replaces the student term with the teacher's own base model π_base, so the reward becomes log π(yt) − log π*_base(yt), isolating capabilities acquired during the teacher's post-training (plus reward centering and an OPD-signal-based conditioning criterion, per the original OPD2 paper). Experiments use Qwen3-1.7B and Qwen3-8B as students and Qwen3-30B-A3B-2507 as teacher, trained for 100 optimization steps on a 100K-problem dataset split 1:1:1 across English/Korean/Japanese (sourced from Nemotron-SFT-Multilingual-v2 and Nemotron-Math-v2), evaluated in both thinking and non-thinking modes on PolyMath, Global-MGSM, HRM8K, and MAWPS.

Key results:

• For Qwen3-1.7B (non-thinking, PolyMath+Global-MGSM avg), OPD2 improves the base model from 55.1→63.6 (English), 40.9→51.9 (Korean), 37.2→52.0 (Japanese); for Qwen3-8B, from 62.8→70.5 (EN), 57.0→64.4 (KO), 57.1→65.0 (JA).
• OPD2 outperforms plain OPD by 3.1–4.0 points in Korean/Japanese for Qwen3-1.7B and 3.1–3.3 points for Qwen3-8B (vs. only 0.7–2.1 points in English), showing the delta signal helps non-English languages at least as much as English.
• OPD2 narrows the English–Korean accuracy gap (thinking mode) on 6 of 7 benchmarks, e.g., Global-MGSM gap reduced from 13.4 to 9.3 points, HRM8K-GSM8K gap from 12.1 to 9.2 points.
• English-only OPD2 training still improves Korean/Japanese non-thinking scores (40.9→52.6 KO, 37.2→53.4 JA), comparable to multilingual training, but the target-language response rate collapses (e.g., Korean thinking-mode final-answer rate drops from 97.6% with multilingual training to 36.1% with English-only training; Japanese drops from 95.2% to 30.8%).

Why it matters / caveats: The results show OPD2's reasoning-transfer benefits generalize cross-lingually, but the authors caution that benchmark accuracy gains from English-only training do not reflect genuine target-language reasoning — models frequently answer in English even when queried in Korean/Japanese — so multilingual training data remains important for preserving target-language generation.

Problem: There is no common, controlled protocol for measuring how well frontier LLMs can automatically optimize an agent "harness" (prompts, tools, control flow, memory, orchestration code) under expensive, stochastic evaluation, since prior methods differ in target agents, seeds, budgets, and disclosure policies, making results incomparable.

Method: HARNESSOPT-BENCH defines an optimizer (an LLM paired with a coding harness) that receives a pinned seed harness H0, graded development/validation feedback (dev reveals per-case traces, validation reveals only aggregate scores), and a fixed evaluation budget (100 calls and 4 full case passes per partition), then edits the harness and nominates a final candidate scored by normalized gain g = (Eθ(H+) − Eθ(H0)) / (1 − Eθ(H0)) on a held-out test partition inaccessible during search. A trusted execution environment (building on VERO) enforces the boundary, meters resource use, and versions candidates for audit. The suite covers 4 downstream tasks (OfficeQA, BrowseComp-Plus, Terminal-Bench, GAIA) with deliberately untuned seed harnesses, and 5 frontier optimizer models (claude-opus-5, claude-sonnet-5, gpt-5.6-sol, gpt-5.6-terra, kimi-k3) are each run under a shared coding harness (opencode) and their native harness (claude-code, codex, kimi-cli), plus two extra harnesses (goose, mini-swe-agent) on GAIA, totaling 111 scored runs.

Key results:

• Changing the optimizer model moves normalized gain by 0.142 on average (holding task/harness fixed), vs. 0.079 for changing the harness (holding task/model fixed) — a ~1.8x larger model effect.
• The best configuration (claude-opus-5) captures up to 0.63 normalized gain on OfficeQA and 0.48 on BrowseComp-Plus, while the weakest (gpt-5.6-terra) is statistically unresolved from zero on BrowseComp-Plus (0.02) and Terminal-Bench (0.04) under opencode.
• Across 5 successive GPT releases on OfficeQA, gain rises monotonically from +0.03 to +0.49; across 5 Claude Opus releases, gain ranges +0.37 to +0.59 (non-monotonic but with a resolvable first-to-last spread).
• Broader exploration of harness "levers" correlates with gain (Spearman ρ = +0.34 to +0.88 across the 4 tasks), while the share of actions spent reading detailed evaluation traces correlates negatively with gain (ρ = −0.31 to −0.64); detailed trace spans were requested only 16 times out of 111 runs.
• Native harnesses show no consistent advantage: across 20 model–task pairs run under both conditions, the shared harness wins 11 and the native harness wins 9 (0 ties), though on GAIA specifically both GPT models score 0.131–0.179 gain-units higher under codex than their best shared harness.

Why it matters / caveats: The benchmark establishes harness optimization as a measurable, discriminative model capability with a trusted evaluation boundary resistant to overfitting on visible scores, but the authors note it is "hack-resistant, not hackproof" (repeated dev/validation feedback could still reward evaluation-specific strategies), the seed harness complexity is not systematically varied, and results are restricted to Python and one pinned target model per task, limiting generalization to other languages/runtimes/architectures.

Problem: Modern Greek is absent from NVIDIA's Nemotron retrieval models and from major multilingual retrieval benchmarks (BEIR, MIRACL), leaving Greek legal/energy/financial/medical RAG applications essentially unmeasured and, as the paper shows, poorly served by off-the-shelf multilingual embedders.

Method: The authors mine a 65,773-pair Greek retrieval training corpus (from 407,053 raw pairs) across five domains using synthetic, style-matched, few-shot-conditioned query generation (Nemotron-3-Ultra-550B and GLM-5.2) with hard negatives mined via cosine-kNN on Qwen3-Embedding-8B; they fully fine-tune Nemotron-3-Embed-1B with an InfoNCE contrastive loss (7 hard negatives/query, GradCache-enabled logical batch of 256 at 4,096-token max length) and separately fine-tune a Nemotron cross-encoder reranker with pointwise binary cross-entropy on 394,579 labeled pairs. For grounded generation, they synthesize 40,000 cite-and-abstain reader examples using a Greek-adapted 27B teacher (Sophea-Titan-1, a LoRA on Qwen3.6-27B) and LoRA-tune (rank 16, α 32, 439M trainable params, 1.37% of model) a Nemotron-3-Nano-30B-A3B mixture-of-experts model (3B active params) as the grounded reader. They also build HERA, a new 4,946-item Greek Wikipedia long-retrieval/RAG benchmark with citations, unanswerable items, and an L1–L5 difficulty ladder.

Key results:

• On specialist Greek domains, BM25 (0.757 macro nDCG@10) beats Qwen3-Embedding-8B (0.680) and the unadapted Nemotron-3-Embed-1B (0.362).
• Fine-tuning the 1B embedder raises macro nDCG@10 from 0.362 to 0.835, a +0.080 margin over BM25 (95% CI [+0.072, +0.089]); on medical retrieval the margin over BM25 narrows to +0.027 (CI touches zero).
• Out-of-domain (HERA Wikipedia retrieval), the adapted 1B beats its own unadapted base by +0.399 nDCG@10, but BM25 (0.6442) beats both the adapted 1B (0.5637) and 0.6B (0.5753) embedders — the in-domain ranking reverses.
• Reciprocal rank fusion of BM25 + adapted 1B embedder gains +0.027 nDCG@10 out-of-domain and +0.013 in-domain over the stronger single component.
• End-to-end two-stage retrieval (embedder+reranker) lifts nDCG@10 from 0.559 to 0.848 (52% relative gain) and Recall@10 from 0.624 to 0.955.
• LoRA-tuned MoE reader: judged answer correctness rises 29.4% → 66.9%; faithfulness rises 25.2% → 84.5%; abstention on unanswerable items improves 1.2% → 30.5% but false-abstention rate also rises 0.4% → 3.4%.
• On a 13-benchmark general-capability sweep, Greek benchmarks improve on 6/8 while English benchmarks regress on 3/5 (e.g., English MMLU −0.098), with mean Δ across all 13 of −0.004.

Why it matters / caveats: The paper's central caution is that adaptation gains are domain-bound — the fine-tuned embedder's win over BM25 does not generalize to general-domain Greek, and BM25 remains a strong or superior baseline across settings, contradicting the common assumption that dense retrieval supersedes lexical search. The authors also explicitly flag several of their own measurement errors (e.g., a first reranker evaluation on 750 queries wrongly concluded the off-the-shelf reranker contributed nothing, corrected upon a larger 2,580-query re-run), retrieval numbers are measured on synthetic (not production) queries and likely overstate real-world lift, and reader-correctness numbers are single-run LLM-judge estimates with a judge sharing model family with the benchmark's generator.

Problem: Training a single generalist vision-language-action (VLA) policy across heterogeneous robot embodiments is hard because existing methods underuse dynamics priors from action-free video and require manual preprocessing to force diverse robot actions into a shared action format.

Method: DyPES-VLA uses a pretrained Qwen3-VL-2B VLM to encode observations, instructions, embodiment metadata, and a set of N=96 learnable query tokens into shared "query states." A future generation head (a SANA-600M image generator) is conditioned on these query states via cross-attention and trained with a rectified-flow objective to predict future frames, teaching the queries to capture object motion, contact, and scene dynamics. An embodiment-specific Mixture-of-Experts (MoE) action head — a 16-layer flow-matching Diffusion Transformer with K=3 experts — shares attention layers across embodiments but routes each sample (via metadata-based static routing) to its own encoder/decoder and FFN expert, generating actions directly in each robot's native action space without common-format alignment. Training has two stages: Stage 1 pretrains the VLM/query tokens/future-generation head on action-free human (EgoDex) and simulation videos (100k steps); Stage 2 jointly optimizes both heads on action-labeled demonstrations from three embodiment families (single-arm, dual-arm, humanoid) for 200k steps, with loss L = L_action + λ_w·L_future (λ_w=0.05).

Key results:

• Single co-trained checkpoint: 98.0% success on LIBERO, 59.25% on RoboCasa-GR1 (29-DoF humanoid), 89.02% on RoboTwin 2.0 (14-DoF dual-arm), each state-of-the-art among single-checkpoint generalists.
• On RoboTwin 2.0, outperforms concurrent generalist Qwen-VLA by 2.37 points average (89.02% vs 86.65%).
• On RoboCasa-GR1, beats best per-benchmark specialist ABot-M0 (58.3%) by 0.95 points and WAM baseline LDA-1B (55.4%) by 3.85 points.
• Real-world: jointly finetuned checkpoint across 3 physical embodiments (FR3, COBOT Magic, G1) averages 75.6% success on 3 tasks, beating GR00T-N1.6 (59.6%) by 16.0 points and ACT (32.4%) by 43.2 points.
• Ablations: removing the future-prediction objective costs 2.4 points on RoboTwin 2.0 and 2.5 on RoboCasa-GR1; replacing MoE head with a shared dense head costs 1.2–2.1 points; removing embodiment metadata costs up to 0.5 points.
• Linear-probe analysis shows future-prediction supervision improves contact-onset AUPRC from 70.8% to 86.3% and contact-release AUPRC from 64.8% to 72.8%.

Why it matters / caveats: The paradigm lets arbitrary embodiments be co-trained without manual action-space alignment, and the linear-probe results give direct evidence that future-prediction supervision embeds genuine dynamics/contact information into the shared representation rather than just improving benchmark scores indirectly. The paper only instantiates and evaluates three embodiment families (single-arm, dual-arm, humanoid), so generalization to further embodiment diversity is not directly demonstrated.

Problem: VLA models typically treat main-view and wrist-view camera observations as parallel/equal inputs, failing to exploit the wrist view's action-proximal role in anticipating fine-grained, contact-sensitive manipulation (e.g., plug insertion).

Method: World-to-Wrist VLA (W2-VLA) uses a Qwen3-VL-4B-Instruct VLM to contextualize K=16 learnable "latent modeling tokens" from current multi-view observations and instructions, forming a fixed-length task-conditioned interface. Conditioned on this interface and wrist history encoded by a frozen V-JEPA 2.1 ViT-L/384 encoder, a 4-layer bidirectional Transformer predictor forecasts future wrist latents (supervised by an L1 loss against a stop-gradient future-latent target), which a Q-Former-style adapter with 32 learnable queries compresses into future-aware wrist context. This context is fused with the VLM's action-conditioning states and fed to a DiT-based flow-matching action head. A separate W2-CoT synthesis pipeline (offline VLM annotator + verification for gripper-state consistency, temporal ordering, release preconditions) produces structured Subtask/Reasoning/Wrist annotations used as auxiliary next-token-prediction supervision during training only (not decoded at inference). Total joint loss combines action flow-matching, CoT next-token prediction, and wrist-latent L1 loss; total model size is ~4.97B parameters.

Key results:

• LIBERO: 98.5% average success (best among 12 compared baselines), with 99.6%/99.8%/99.2%/95.2% on Spatial/Object/Goal/Long suites, exceeding the strongest baseline by 1.3 points.
• RoboTwin 2.0: 60.71% success under clean (Easy) setting (vs. 52.92% for UP-VLA, +7.79 points) and 18.21% under domain-randomized (Hard) setting (vs. 16.34% for π0, +1.87 points).
• Real-world (3 CoBoT Magic tasks): standard-condition average success 70.00%, beating VLA-JEPA by 15.56 points and π0 by 28.89 points; OOD average success 52.22%, beating VLA-JEPA by 14.44 points (largest margin on Bimanual Plug Insertion: 33.33% vs 10.00%).
• Inference: generates a 16-step action chunk in 183 ms (87.43 Hz), over an order of magnitude faster than explicit CoT decoding (>1.5 s/chunk) with comparable or higher accuracy (98.5% vs 97.6–98.1%).
• Ablations: removing the wrist predictor drops LIBERO average from 98.5% to 97.5% (largest drop on Long suite, 95.2%→93.6%); removing W2-CoT supervision drops it to 98.0%; predicting wrist-view futures only outperforms predicting main-view only (97.7%) or both jointly (98.0%).

Why it matters / caveats: The wrist-future-prediction target is shown to matter more than jointly predicting the main view, and the fixed-length latent interface avoids the inference-time cost of explicit chain-of-thought decoding while retaining most of its benefit — enabling real-time (>80 Hz) deployment. Real-world evaluation is limited to a single robot platform (CoBoT Magic) and three tasks, and success rates in the hardest RoboTwin 2.0 setting (18.21%) remain low in absolute terms.

Problem: Video object removal must eliminate not just the target object but also its induced effects (shadows, reflections, ripples, smoke, etc.), and existing methods learn object-effect correspondences only implicitly from a limited set of predefined effect categories, generalizing poorly to compositional, spatially detached, or long-tail real-world effects.

Method: EffectLearner combines a VLM-based Object-Effect Reasoner (Qwen2.5-VL-3B-Instruct) with a DiT-based Video Eraser (Wan2.2-TI2V-5B, initialized from Kiwi-Edit). Given a target-highlighted video (alpha-blended mask overlay with a colored boundary) and a structured effect-analysis prompt, the VLM produces hidden states that learnable effect queries (Q-Former-style) aggregate into compact effect-aware context tokens, which replace the text condition in the DiT's cross-attention layers; the source video latent and downsampled object mask condition the DiT via channel concatenation with the noisy latent. Training uses a region-weighted flow-matching loss that up-weights object and effect regions (effect regions found by thresholding source-target latent differences), plus motion-aware mask guidance (temporal union of per-frame masks to counter VAE temporal compression) and a motion-consistency loss that matches motion-compensated latent residuals across adjacent frames via backward-warped pseudo-flow. The authors also construct EffectWorld, an Unreal-Engine-rendered paired dataset (11,092 valid triplets) covering compositional, weakly correlated, and long-tail physical effects, and train with a progressive curriculum moving from basic object removal to complex-effect data.

Key results:

• On EffectWorld-Eval, EffectLearner achieves best or near-best across all metrics: PSNR 29.521, SSIM 0.934, LPIPS 0.044, MAE 6.839, FVD 69.184 — clearly ahead of ROSE (FVD 887.541) and EffectErase (FVD 238.521).
• On ROSE-Bench, EffectLearner attains the best LPIPS (0.054) and substantially lower FVD (81.866) than ROSE (803.722) and EffectErase (135.117), despite slightly lower PSNR (28.973 vs ROSE's 30.468).
• On EffectWorld-Wild (VBench-style, no-reference), EffectLearner achieves the best total score (0.772, tied with VACE) with the strongest Background Consistency (0.958) and Dynamic Degree (0.400).
• Ablations on EffectWorld-Eval: removing VLM guidance drops PSNR from 29.521 to 27.737 and raises FVD to 118.236; removing the motion-consistency loss causes the largest degradation (PSNR 24.703, LPIPS 0.187, FVD 294.778), showing spatial supervision alone cannot ensure temporal coherence.

Why it matters / caveats: The results indicate that explicit semantic reasoning about object-effect relations (rather than category-specific supervision) improves generalization to compositional and long-tail real-world effects, and the ablations isolate that motion-consistency supervision is the single largest contributor to output quality. Evaluation on EffectWorld-Eval and EffectWorld-Wild uses benchmarks the same authors constructed, and EffectWorld-Wild's non-DAVIS masks are only semi-automatically annotated (detector + SAM2 + manual correction) rather than fully independent ground truth.

Problem: Existing benchmarks for natural-language data analytics fail to unify realistic workspace-scale evidence discovery (across databases, files, long documents, and video), the requirement to return a complete tabular result, and deterministic evaluation, leaving no benchmark that stresses all three simultaneously.

Method: The authors build DataSpace, a 410-task, cross-language (Chinese/English) benchmark spanning CSV, JSON, SQLite, Markdown, PDF, and video (7,439 artifacts, ~15 GB), constructed via DataSpace-Builder — a four-stage pipeline (Cross-Language Transformation, Constraint-Aware Relational Sampling, Modality Routing & Artifact Rendering, and Human Review & Task Repair) that transforms EHRSQL and BULL Text-to-SQL instances into heterogeneous workspaces with expert-reviewed gold tables. Evaluation uses a deterministic evaluator performing header-invariant column alignment (searching permutations for a valid column mapping), type/precision-aware value normalization, and order-aware or order-agnostic row comparison. Agents are tested via a fixed ReAct-style DataSpace-Agent harness across six frontier multimodal backbones, and via five different agent harnesses with one backbone fixed.

Key results:

• Best backbone (Grok 4.5, DataSpace-Agent fixed) reaches 66.34% Task Accuracy (272/410); GPT-5.6 Sol reaches 64.63%; weakest (MiniMax M3) reaches 28.54%.
• Harness choice creates a 15.36-point spread with MiMo-V2.5 backbone fixed: Grok Build 46.34% vs. Smolagents 30.98%.
• Oracle union across six backbones solves 334/410 tasks (81.46%), while 76 tasks (18.5%) are solved by none.
• GPT-5.6 Sol trails Grok 4.5 by only 1.71 points while using 74.2% fewer tokens, 50.3% fewer tool actions, and 39.2% less latency; MiMo-V2.5 is cheapest at $0.011/task vs. $0.169 (Grok) and $0.200 (GPT).
• Multimodal (cross-modality) tasks reduce accuracy for every backbone by 1.8–14.0 points; join-requiring tasks reduce accuracy by 9.7–19.8 points across all backbones.

Why it matters / caveats: DataSpace remains far from saturated (best accuracy ~66%), showing current data agents still struggle with cross-modal evidence integration and relational joins over realistic, heterogeneous, multilingual workspaces; the benchmark is derived from only two source Text-to-SQL corpora (financial/macroeconomic/healthcare domains), which may limit topical generality.

Problem: Existing multimodal large language models (MLLMs) for 3D scene understanding fuse a fixed set of modalities (RGB, depth, BEV, point cloud, voxel) for every query, even though different question types (e.g., color vs. shape vs. location) genuinely need different modalities, causing irrelevant modalities to introduce semantic noise and wasted computation.

Method: SmartMage, built on Qwen3-VL-8B-Instruct, has two components. (1) SMART (Semantic-guided Modality Adaptive RouTing) is a global modality scheduler that combines three signals — a Semantic Prior Estimator (a linear head over the text embedding predicting a modality relevance distribution), a Semantic Similarity Scorer (cross-attention between modality-specific query vectors and each modality's features, then cosine similarity with the global instruction embedding), and a Modality Quality Evaluator (a linear head over per-modality activation strength/sparsity/stability statistics) — to select RGB (fixed primary) plus a query-relevant subset of complementary modalities. (2) MAGE (Modality-Aware Gating Expert) inserts sparse MoE layers (8 experts, top-2 routing, initialized from pretrained FFNs) into LLM layers 8/12/16/20/24/28; a Modality-aware Expert Speculation (MES) module predicts a per-token modality distribution and, via a learnable modality-expert affinity matrix, biases the MoE router toward modality-consistent expert assignment. Training is end-to-end with frozen visual encoders and frozen base LLM parameters (only adapters and router/expert branches trainable), using a combined loss of cross-entropy plus semantic-alignment losses (L_sem, L_dis) and expert-assignment losses (L_ma, L_ec, L_bal).

Key results:

• On five standard 3D benchmarks: 32.6 EM@1 on ScanQA and 66.8 EM@1 on SQA3D (vs. prior SOTA Ross3D's 30.8 and 63.0); 88.7 CIDEr@0.5 on Scan2Cap (vs. Video-3D LLM's 83.8, +4.9); 59.5 Acc@0.5 on ScanRefer (vs. Ross3D's 54.4, +5.1) and 60.7 F1@0.5 on Multi3DRefer (vs. Ross3D's 54.3, +6.4).
• Ablation on modality selection: fixed all-5-modality fusion (setting 10) scores 30.7 EM@1 on ScanQA and 57.2 F1@0.5 on Multi3DRefer, while adaptive selection reaches 32.6 and 60.7 respectively — outperforming every fixed combination tested, including combinations using fewer modalities that sometimes beat combinations using more (e.g., setting 8 beats settings 9 and 10 despite fewer modalities).
• Removing SSS (Semantic Similarity Scorer) causes the largest SMART-component drop (ScanQA EM@1 29.8→27.1; Multi3DRefer F1@0.5 56.2→52.2); removing MES causes the largest MAGE-component drop (SQA3D EM@1 64.5→63.1).
• Training efficiency: 47.44 s/iteration vs. 91.61 s (Video-3D LLM) and 125.2 s (Ross3D) — a 2.0×–2.6× speedup — while End-to-End inference latency (538.2 ms) remains comparable to baselines (527.0–551.6 ms).
• On RGB-only video benchmarks (zero-shot, no 3D input), SmartMage remains competitive: 27.8 average on VSI-Bench, 31.8/30.0/30.0/28.3 on VSI-SUPER-Recall at 10/30/60/240 min, and 30.3 average on MMSI-Bench, generally ahead of other 3D-scene-understanding methods evaluated on RGB-only input.

Why it matters / caveats: The ScanFacet diagnostic benchmark (built by the authors, reorganizing ScanQA/SQA3D into eight semantic facets) provides direct evidence that different question types indeed prefer different modality combinations (e.g., color/material favor RGB or RGB-Depth; spatial/counting favor Depth/Voxel), supporting the paper's core motivating claim; the largest facet-level gains are +27.1 and +15.9 CIDEr on material and color understanding. The comparisons are against the authors' own re-implementations/adaptations on a self-constructed ScanNet-derived training corpus, and RGB-only 2D-specialist video models still outperform SmartMage on some benchmarks (e.g., VSI-Bench), which the paper attributes to their longer temporal context and video-specific training.

Problem: Executable validation of synthesized terminal-agent training tasks confirms feasibility but does not reveal whether a task is appropriately challenging (neither trivial nor effectively unsolvable) for a given solver, i.e., it says nothing about learnability.

Method: CalibForge places an authoring agent (DeepSeek-V4-Pro) in a draft sandbox to research clues via web search and construct candidate terminal tasks (instruction, Dockerfile/environment, verification tests), which pass two-stage validation (structural checks + self-solving) before entering an adversarial author–solver calibration loop (up to 50 rounds). Two calibration strategies are used: Multi-Solver Calibration, which retains a candidate only if outcomes disagree across a pool of solver models (DeepSeek-V4-Flash, GLM-5, Kimi K2.5) — at least one pass and one fail; and Contrastive Solver Calibration, which retains a candidate only when a designated stronger solver (DeepSeek-V4-Pro) passes and a designated weaker solver (DeepSeek-V4-Flash) fails. Retained tasks (5,431 total) are converted to Harbor-style task instances, SFT trajectories are distilled via DeepSeek-V4-Pro under a minimal bash/file-edit/finish scaffold, and used to fully fine-tune Qwen3-30B-A3B-Instruct and Qwen3.5-35B-A3B for 10 epochs.

Key results:

• CalibForge-30B-A3B and CalibForge-35B-A3B achieve 32.58% and 47.57% on Terminal-Bench 2.0, beating the strongest re-distilled baselines (TerminalTraj/TermiGen) by 6.36 and 6.75 percentage points respectively.
• OOD transfer: 30B model improves 27.68 points on SWE-bench Pro Resolved Rate (3.26%→30.94%) and 30.04 points on Doc2Repo Pass Rate (5.94%→35.98%) over the base model; 35B model improves 3.03 and 3.85 points respectively.
• Ablation (matched 1,300-task setting): No-Solver 22.47% → Single-Solver 24.34% (+1.87) → Multi-Solver 29.21% (+6.74) → Contrastive-Solver 31.09% (+8.62) on Terminal-Bench 2.0.
• Under contrastive calibration, only 19% of candidates initially satisfy the target strong-pass/weak-fail relation despite having passed structural validation and self-solving; after revision/re-probing, 96% ultimately satisfy it.
• Resulting task collection spans 3,885 distinct capability tags (median 5/task) and its largest category (software engineering, 25.5%) is far less dominant than in prior datasets (e.g., SETA-Env is 74.6% system administration).

Why it matters / caveats: The gains come specifically from the calibration criteria rather than from more training data — multi-solver calibration actually produces fewer retained SFT trajectories than the no-solver variant yet improves accuracy substantially. Results are demonstrated only on two Qwen backbones and rely on a fixed set of solver models (DeepSeek-family, GLM-5, Kimi K2.5) for calibration, and the authors note some hard candidates require sustained probing (up to 20 rounds) to become learnable.

Problem: Deep networks can retrieve short segments of perceived speech from MEG with high accuracy, but their learned weights cannot be mapped to standard electrophysiological notions (cortical location, rhythm, time course), and it is unclear which properties of the speech stimulus actually drive the decoder's decisions.

Method: Building on the factorized spatial-temporal interpretability framework of Petrosyan et al., the authors replace the 2D Fourier spatial-attention layer of a prior MEG-to-audio contrastive decoder (Défossez et al.) with a spherical-harmonic-parameterized 3D spatial attention layer (L=24, 576 basis functions per virtual channel, J=270 virtual channels), reduce the subject-specific branch representation from 270 to K=25 branches, and add a trainable 150ms depthwise temporal filter per branch, so each branch is jointly matched to a neural source in space and time; ocular/cardiac ICA components are removed before training to prevent shortcut learning. The model is trained with a CLIP-style MEG-to-wav2vec-2.0 contrastive objective (AdamW, batch 100, up to 50 epochs) on the MEG-MASC dataset. Branch weights are mapped to cortical source space via Minimum Norm Estimate and RAP-MUSIC, and stimulus-feature relevance is probed via paired MEG occlusion — replacing feature-marked speech intervals with matched real-MEG donors from feature-present vs. feature-absent intervals and comparing retrieval rank shifts (participant sign-flip permutation test, max-T FWER correction).

Key results:

• 39.75 ± 0.34% Top-1 accuracy (70.4% Top-10) among 1005 candidates, averaged over six trained solutions, with the full decoder using 486,619 trainable parameters (~20× fewer than Défossez et al.'s decoder).
• Within the authors' own ablation grid, a K=270, five-block configuration is 14.8× larger than the main model yet scores 3.60 percentage points lower in Top-1 accuracy.
• Paired occlusion analysis: 15 of 19 tested stimulus features show significant positive effects (pFWER<0.05); largest effects are silence (Δrank=75.62), high loudness (60.77), vowels (38.70), and strong acoustic onset (36.19), all pFWER<10⁻⁴ and positive in all 27 participants.
• Randomly ordered word lists show a significant negative effect (Δrank=−17.77): narrative MEG substituted into word-list intervals improves retrieval, unlike all other tested features.
• The wav2vec target can be compressed from 768 to ~12 learned feature dimensions (via a task-trained linear projection) with retrieval accuracy essentially unchanged, whereas unsupervised PCA degrades faster at low dimensions, and temporal compression (e.g., mean/max pooling to a single vector) collapses accuracy to near-chance.

Why it matters / caveats: The work demonstrates that physically/physiologically constrained decoders can match or approach black-box decoding accuracy while remaining interpretable, recovering cortical sources consistent with the canonical speech-perception network (bilateral superior temporal gyrus, dorsal-stream sites, left frontal operculum) and a left-lateralized higher-frequency (~13.3 Hz) component absent on the right. The authors note their accuracy figures are not directly comparable to Défossez et al. because their test segments are not word-onset-aligned (a harder setting), and effect-magnitude ordering across occlusion features is confounded by differing replaced-interval durations, so it should be read as identifying which features matter rather than precisely how much each contributes.

Problem: End-to-end autoregressive document parsers serialize all page-layout and regional content into one sequence, forcing spatially independent regions onto a single decoding path whose length grows with total page content, making inference slow; crop-based two-stage parsers get region-level parallelism but lose full-page context and require repeated visual encoding of crops.

Method: PaDoc is a single 2.1B-parameter MLLM (initialized from Qwen3-VL-2B) that models document parsing as a prefix-conditioned factorization: under a region-sufficiency conditional-independence assumption, each region's content Yk depends only on the shared page image X and the layout prefix B≤k (not on future layout or sibling content), so the layout stream and per-region content branches can be decoded concurrently rather than sequentially. Training uses packed variable-length ("tree-varlen") ancestor attention — realized as a set of variable-length causal attention problems rather than a dense mask — with a standard next-token objective over two stages: continual pre-training (1 epoch, ~11M examples) followed by supervised fine-tuning (0.5M curated examples), using global batch size 256, peak LR 5×10⁻⁶, bf16/FSDP across 128 A800 GPUs. At inference, masked parallel decoding forks branch-isolated content streams that are served as concurrent vLLM requests with cache-resident shared-prefix (image+layout) reuse.

Key results:

• On OmniDocBench Full, PaDoc attains Overall layout F1 of 91.1 (IoU/F1 81.4/91.1), with the highest Overall precision (93.3) among end-to-end parsers and leads all four table metrics (92.7 IoU, 97.0 F1).
• End-to-end parsing Overall score of 94.24, exceeding Qianfan-OCR (93.90) by 0.34 points and the strongest general VLM evaluated (Ovis2.6-30B-A3B, 93.62) by 0.62 points; best end-to-end Text Edit (0.038) and Formula CDM (95.59).
• On a 384-page subset with one A800 GPU across 5 concurrency levels (16–256), PaDoc improves valid-page throughput by 67.4–118% and reduces P95 latency by 39.2–54.9% relative to a same-backbone Sequential SFT baseline; throughput peaks at 1.722 pages/sec/GPU at concurrency 64.
• Despite its 2.1B size, PaDoc surpasses the 1.0B HunyuanOCR-1.5 in both throughput and P95 latency at every tested concurrency, and substantially outpaces the 0.7B MonkeyOCRv2-B-Parsing.

Why it matters / caveats: PaDoc shows that layout-grounded parallelism can be added within a single MLLM (no separate detector/recognizer heads, no auxiliary losses) to cut serving cost without sacrificing parsing quality, closing much of the gap to specialized two-stage systems. The evaluated efficiency numbers reflect a specific vLLM serving realization (request-level, not a physically materialized ancestor mask), so the reported speedups include real-world overheads like scheduling and branch-length imbalance rather than an idealized theoretical bound.

Problem: Yiddish is severely underserved by LLMs because existing multilingual pretraining corpora and benchmarks are poor proxies for the language — containing substantial noisy, machine-translated, or misclassified (e.g., mislabeled Hebrew) text — and no systematic multi-task evaluation resource previously existed for Yiddish.

Method: The authors first audit the Yiddish split of mC4 (manual/automatic domain and language-ID inspection of URLs and sampled pages), then build Oytser, a new ~915M-word (5.28B-token) Yiddish pretraining corpus combining nine web-native (news, forums, Wikipedia, Bible translation) and literary sources, the largest being the OCR'd Yiddish Book Center corpus (12.3K books, ~4.23B tokens via Jochre 3 OCR). They also construct Kashes, a 9-task benchmark (translation, POS tagging, dependency parsing, transliteration, lemmatization, 3 NER datasets, and 3 machine-translated language-understanding tasks from Aya), including a new natively-authored Yiddish–English translation set (Kashes-mt, 5,287 sentence pairs) built via document matching, LaBSE-based sentence alignment (SentAlign), and quality/deduplication filtering from bilingual Forverts and In geveb publications. MameLoshnLM is produced via continued pretraining (causal LM objective, 1 epoch, ~5.7B tokens, bfloat16, 8-bit AdamW, LR 2×10⁻⁵) of Llama-3.1-8B on a Yiddish-dominant (72% words) + English (CC100) data mixture on a single H200 GPU (~207 GPU-hours).

Key results:

• mC4 Yiddish-split audit: only 42.2% of documents are genuine high-quality native Yiddish; 29.8% are machine-translated and 21.9% are misidentified Hebrew.
• MameLoshnLM achieves the best average score (62.6) across 14 Kashes evaluations (5-shot), ahead of Gemma-2 9B (57.0), Llama 3.1 8B (56.8), Qwen3 8B (54.7), and BLOOMZ 7B (20.9); it leads on POS tagging, dependency parsing, transliteration, and EHRI/newNLP NER.
• English→Yiddish translation: MameLoshnLM beats the closest competitor by more than 11 COMET points (78.5 vs. 66.7 for Gemma-2 on FLORES+; 75.3 vs. 59.9 on Kashes-mt).
• Loshn-koydesh (Hebrew/Aramaic-derived) vocabulary probe: MameLoshnLM produces LK content words at a 4.7% rate vs. 1.6% for Llama 3.1 8B (gold reference rate 6.2%, p<10⁻²²⁹); sentence-level LK match rate 52.4% vs. 16.0%.
• Morphology (lemmatization change accuracy): on irregular ge- participles, MameLoshnLM reaches 50.8% vs. 5.1% for Llama (p<10⁻²⁹); on regular "האָבן" (to have) forms, both models tie at 84.7% (not significant).

Why it matters / caveats: The results show that noisy web-scale multilingual data can yield superficially fluent but non-native output for low-resource languages, and that targeted corpus curation plus continued pretraining can substantially close this gap — a template potentially applicable to other historically rich, digitally underrepresented languages. The authors note that reallocating training budget to historically related languages (Hebrew, German) instead of English gave only limited and inconsistent gains, and some Kashes tasks (PAWS-Wiki, PIQA) rely on machine-translated Aya benchmarks where MameLoshnLM underperforms Qwen3 8B.

Problem: Existing video models expose text-to-video generation, reference conditioning, and editing as separate operations over fixed inputs, so no single model can compose these interactively over an evolving multi-shot visual history without denoising cost growing as the history grows.

Method: ContextMaster (initialized from Wan2.1-T2V-1.3B) uses role-aware rotary coordinates (RoPE phase offsets keyed by role/shot index) to unify reference, history, source, and target tokens in one sequence; a cacheable fixed-budget context mechanism prefills reusable clean-context KV states once per round and restricts each target query to an absolute budget of B context blocks via block-sparse attention, with a "ConstraintSink" reserving mandatory reference and temporally-aligned source blocks while the remainder is routed by content similarity. The deployment model is trained with two-stage privileged context distillation: (1) privileged consistency distillation transfers a dense full-context teacher's denoising behavior to the sparse few-step student via a two-endpoint consistency loss, then (2) distribution matching distillation (DMD) refines deployment-matched rollouts using a frozen dense real-score model and a trainable fake-score model. Deployment uses 4 denoising steps and a 6-frame-equivalent (6 FE) active-read budget with a TileLang block-sparse kernel.

Key results:

• Reaches 16.74 FPS on a single H200 GPU across five-shot T2MV/R2MV/V2MV runs.
• T2MV: Inter-Shot consistency 0.836 vs. 0.808 for the next-best baseline (ShotStream); task fulfillment (TF) 4.17 vs. 4.03 (ShotStream).
• R2MV vs. Phantom: TF 4.04 vs. 3.68; Inter-Shot 0.749 vs. 0.638; FPS 16.74 vs. 0.452.
• V2MV: leads or ties on 8 of 9 reported metrics vs. VideoCoF/LucyEdit/StreamEdit/LiveEdit (e.g., TF 4.20 vs. 3.80 best baseline).
• User study: highest instruction-following (IF) and cross-shot consistency (CC) scores among primitive tasks vs. strongest baselines.
• Ablations: removing role-aware RoPE or ConstraintSink drops macro-averaged Inter-Shot from 0.779 to 0.748/0.742; PCD→DMD training beats vanilla sparse training or either stage alone on every reported metric (e.g., TF 4.17 vs. 3.92 vanilla).

Why it matters / caveats: The architecture bounds active-context reads independently of history length, enabling interactive multi-shot workflows (including a tool-augmented "Director" agent) at real-time-ish speed. The authors note a limitation: throughput still decreases by roughly 0.4 FPS per additional shot because the context branch performs bidirectional prefill over the full accumulated history to build its KV cache.

Problem: Vision encoders are known to be sensitive to imperceptible pixel-level metadata (e.g., JPEG compression, camera parameters), but the cause of this sensitivity and how to mitigate it were unclear.

Method: The authors hypothesize metadata sensitivity arises from metadata-semantics correlations in pretraining data, and test this with diagnostic metrics from prior work (Ramos et al.): Metadata Prediction (MP, linear-probe accuracy for processing/acquisition metadata), Semantic Prediction (SP, k-NN accuracy), and Semantic Prediction Distraction (SPD, accuracy/recall gap when metadata is manipulated to align with positives vs. negatives). They measure natural correlations via Cramér's V on ImageNet1k/21k and topic-modeled Re-LAION-2B, then synthetically control correlation strength (parameter pi, 0–100) by reassigning JPEG compression labels per class and retraining ResNet50/ViT models, and analogously create stronger/weaker/baseline Re-LAION-2B subsets for CLIP-loss training. For mitigation, they test (a) training-time data augmentations (RandomAugment variants, color jitter, grayscale, blur) and (b) a post-hoc adversarial linear-layer mitigation trained via a two-step adversarial procedure (frozen mitigation layer trains a metadata classifier, then the mitigation layer is trained to fool it while preserving/decorrelating features).

Key results:

• Cramér's V association between JPEG metadata and semantic labels: 0.047 (IN1k) vs. 0.067 (IN21k), with IN21k-trained models showing consistently higher MP and SPD.
• Synthetic correlation control (ResNet50 on IN1k): increasing pi from 0 to 100 raises JPEG MP accuracy from near-chance toward much higher, and also raises MP for unrelated attributes (resizing, camera make, aperture), showing generalized shortcut learning.
• LAION acquisition-metadata subsets (Cramér's V 0.396 "stronger" vs. 0.166 "weaker"): stronger correlation yields higher SPD and lower semantic prediction (SP) accuracy.
• Adversarial mitigation guided only by JPEG labels also reduces sensitivity to unseen metadata types (resizing, camera make/model) across CLIP, ConvNeXt, SigLIP, ViT, DINO, and DINOv2 models, while SP is maintained or slightly improved.
• Generated-image detection: ResNet50 trained with pi=100% correlation achieves average detection accuracy 60.4 vs. 57.4 for pi=none across GAN/deepfake/low-level-vision/diffusion detectors (Table 1).
• OOD gains from mitigation-during-training augmentations on ResNet50/IN1k: ImageNet-C 44.2→45.0, ImageNet-R 34.5→37.7, ImageNet-Sketch 26.4→30.0.

Why it matters / caveats: Metadata sensitivity is double-edged: it explains why some encoders (e.g., CLIP) are strong generated-image detectors, but it also causes fragility under metadata distribution shift; mitigation improves OOD generalization most for the more sensitive models (CLIP-family, supervised ConvNeXt) while sometimes hurting already low-sensitivity models like DINOv2. The authors note it remains open how strongly this contamination affects downstream tasks and standard benchmarks.

Problem: Current video world models entangle world dynamics with view-dependent visual latents, so multiplayer simulation requires a separate recurrent history per camera, causing redundant compute, cross-view inconsistency, and simulation cost that scales with the number of viewers.

Method: MASS separates a learned Logic Engine from a learned Rendering Engine, mirroring authoritative-server game architectures. A game schema declares typed entities (kinds, fields, counts); at each tick a World State Tokenizer converts the typed state into short per-record token sequences, and a single decoder-only Transformer (width 256, 6 layers, 8 attention heads, tied embeddings; 5.66M parameters in the matched-benchmark config) autoregressively predicts each record's next-state tokens under schema-derived masks, with a deterministic assembler reconstructing the full next state (Kg=5,121 records for the 1,024-player Snake instantiation). A separate geometry-aware residual U-Net Rendering Engine (3 downsampling stages, 5 dilated residual blocks, 3 skip-connected upsampling stages) renders any requested camera view from a camera-local projection of the predicted typed state, independent of the number of cameras. Renderer training uses object-weighted Charbonnier reconstruction, multiscale L1, image-gradient alignment, SSIM, and color-saturation losses (no perceptual/adversarial loss).

Key results:

• On the matched multiplayer Snake benchmark: MASS achieves LPIPS 0.098 vs. 0.123 (B-UN dense-state baseline) and 0.277 (MultiWorld).
• Parser-based state recovery reaches 0.764 vs. 0.128 for the best video-based baseline (B-PV), a >5x gain.
• Cross-view disagreement (X-view) is 0.000 for MASS vs. 0.984–1.000 for MultiWorld/B-PV/B-SL.
• Direct logic (state-only) evaluation: MASS reaches 99.1% head-position accuracy at H=1 and retains 90.2% at H=32, with 0% structural contradiction, vs. ≤2.7% position accuracy and 100% contradiction rate for three dense-state diagnostics (independent-head CNN, RSSM, joint U-Net).
• The same architecture scales to 1,024 concurrent simulated player entities for 10,000 recurrent ticks, with per-tick simulation cost independent of the number of rendered cameras.
• Held-out reconstruction across 8 games: PSNR ranges 23.72 dB (Tron) to 40.24 dB (Frogger), 5 of 8 games above 32 dB.

Why it matters / caveats: Explicit typed state lets errors be localized to dynamics vs. rendering before any frame is generated, and B-UN's ablation shows that low LPIPS (visual similarity) does not guarantee a recoverable world state (it has 0% parser recovery despite competitive LPIPS). The authors note extending the approach to 3D environments and richer entity interactions remains future work, and world recoverability still degrades over long rollouts (e.g., MASS position recall falls from 1.000 at tick 1 to 0.224 at tick 128 in the extended results).

Problem: The classical continual learning (CL) taxonomy — parameter-centric updates via off-policy gradient optimization during a dedicated training stage — no longer captures emerging LLM/agentic-AI paradigms such as on-policy post-training, test-time training, and external memory/skill/protocol systems.

Method: This is a survey/position paper, not an empirical algorithm paper. The authors propose a tri-axial taxonomy for characterizing continual learning methods: "When" (pre-training, post-training, inference-time), "Where" (model parameters vs. external "harness" carriers — memory, skills, protocols), and "How" (off-policy gradient updates, on-policy updates, and beyond-gradient mechanisms such as model merging, zeroth-order optimization, heuristic learning, and prompt evolution). They place classical CL as the reference point (post-training, parameters, off-policy) and use the framework to review and organize representative methods (e.g., continual pretraining recipes, LoRA-based parameter-efficient CL, memory systems like MemoryBank/A-MEM, skill libraries like Voyager/SkillRL, protocol evolution like Promptbreeder, on-policy methods like TTRL, and gradient-free methods like Task Arithmetic/ZeroFlow) into this common space.

Key results: Not stated — this is a survey paper organizing existing methods along the proposed framework rather than reporting new experiments, benchmarks, or numerical results of its own.

Why it matters / caveats: The framework offers a way to unify disparate recent directions (harness-based agent memory/skills/protocols, inference-time adaptation, gradient-free updates) under one lens and to identify underexplored regions (e.g., protocols as a CL carrier are noted as "at an early stage" relative to memory and skills). The authors themselves caution that the method placements in their figures are illustrative and not from an exhaustive literature-enumeration protocol, so density in a region should not be read as a statistical estimate of research activity, nor should sparse regions be read as evidence of no relevant work.

Problem: Adapting multilingual text embedding models with a single training objective across heterogeneous task families (translation, retrieval, classification, pair-classification) produces conflicting training signals, since these tasks require fundamentally different optimization strategies.

Method: Task-Conditional Flow Matching (TCFM) applies Flow Matching (a velocity-field regression objective, with the target velocity optionally blended with a frozen teacher's translation direction) only to translation-style sentence pairs, while retrieval, classification, and pair-classification are trained with a multi-positive InfoNCE contrastive loss (symmetric for translation, asymmetric for retrieval); a bounded transport objective lets the learned flow field influence embeddings without large shifts. Additional losses include pointwise and relational (Gram-matrix) teacher-preservation terms, a cosine-margin hard-negative repair loss (activated only for task families with explicit hard negatives), and a SimCSE-style monolingual consistency loss. Training follows a three-stage curriculum: Stage 1 cross-lingual alignment (translation only, heavy Flow Matching/teacher weighting), Stage 2 multi-task semantic adaptation (adds classification/NLI with a ~23% translation replay buffer), and Stage 3 retrieval adaptation plus monolingual regularization (~41% replay). The velocity network is a two-layer MLP with SiLU activations, LayerNorm, and sinusoidal timestep embeddings, conditioned on a frozen teacher's instruction embedding.

Key results:

• On Indic MTEB v3 (7 task categories, 25 languages), Harrier-0.6B improves from 65.87 to 69.46 average score (+3.59 points; +5.45% relative), with Clustering improving by +21.03 points.
• Qwen3-Embedding-8B (adapted via LoRA) improves from 72.89 to 74.87 (+1.98 points; the paper also reports +1.97/+2.72% in text), gaining on 6 of 7 categories while Retrieval dips slightly (-0.28).
• Cross-lingual gains: IN22ConvBitextMining +3.30 and IndicCrosslingualSTS +2.24 (Harrier-0.6B).
• Ablation (Harrier-0.6B): full TCFM (Flow + Contrastive + Teacher) reaches +3.59 macro average vs. +1.96 for contrastive-only and +2.81 for contrastive+teacher without flow matching; removing teacher preservation while keeping Flow+Contrastive drops Retrieval by -12.48 and STS by -8.99 points.
• Restricting Flow Matching to translation only outperforms applying it to all tasks (+1.67 vs. +1.63 macro average).
• A few tasks regress after adaptation, e.g., SanskritShlokasClassification -5.31, XQuADRetrieval -0.23 (Harrier-0.6B).

Why it matters / caveats: The results support routing loss objectives by task semantics rather than using one contrastive loss for everything, and gains generalize across encoder scales (270M–8B) and both bidirectional-encoder and decoder-only (Qwen) architectures. The authors note the evaluation covers only 22 Indic languages (generalization to other language families untested), the claim that Flow Matching improves embedding-space uniformity is empirical rather than formally/geometrically verified, TCFM adds computational overhead (frozen teacher, velocity network, multi-stage curriculum), and it depends on high-quality parallel translation data, limiting applicability to zero-resource languages.