Multi-token prediction (MTP) accelerates language model decoding by drafting several future tokens in one forward pass, but the speculation horizon must be fixed before drafting and the best length varies with context. This paper frames horizon selection as an online survival problem: it predicts per-position rejection hazards to forecast how long a draft prefix will survive exact verification, then adaptively picks the horizon maximizing expected tokens per second. The resulting controller, AMTP, requires no retraining, leaves the target verifier unchanged, and comes with a sublinear throughput-regret bound against the best horizon at each visited context.

DuoCook evaluates whether a single vision-language model can centrally coordinate two cooking agents in Unreal Engine kitchens inspired by Overcooked. The benchmark spans 20 maps across five categories—basic, shared equipment, dynamic terrain, combined dynamic/shared, and separated workspaces—requiring the model to assign high-level actions from top-down visual observations while the engine handles low-level navigation. A memory ablation tests the value of a shared, self-authored belief notebook. Performance is far from saturated even with heavy scaffolding, suggesting current VLMs remain poor at real-time centralized coordination, though the task interface provides substantial structured assistance.

Regulated institutions must produce technical evidence for LLM deployments under the EU AI Act, yet existing benchmark suites are engineer-facing command-line tools while accountable reviewers are non-technical managers. VERA is an open-source, self-hosted dashboard that closes this workflow gap by letting reviewers select a model, run public benchmarks, and read requirement-level scores—with decomposable evidence and run-level energy consumption—in a browser. Its core contribution is wrapping raw measurement into a review and filing process designed for institutional approval.

The paper tests whether open LLMs that a hospital could self-host are ready for clinical use, evaluating four models (8–27B parameters) on the Spanish MIR residency exam under its real scoring and ranking rules. The authors go far beyond headline accuracy to probe calibration, the value of abstention under a wrong-answer penalty, the soundness of clinical reasoning, and whether vision models actually use the images they receive. Their headline message is that a model can rank high enough to choose almost any specialty while remaining overconfident, correct for the wrong reasons, and blind to the clinical plates it is shown.

SABLE tackles expensive optimisation in the high-dimensional latent spaces of pretrained generative models — a regime where latent dimension $D$ can reach $40{,}500$ while the evaluation budget $N$ is orders of magnitude smaller. The core idea is to use a Bayesian affine surrogate on the characteristic-radius sphere, decoupling acquisition into a data-sized in-span component and a scalar off-span norm that is prescribed by a finite-horizon schedule rather than left to expected improvement. Together with a fast recompression scheme that bounds model-refitting cost to $\mathcal{O}(M^{2}D+M^{3})$, the method achieves tens-of-millisecond decisions and strong sample efficiency on synthetic and scientific benchmarks.

Tool selection across corpora of tens of thousands of APIs is a critical bottleneck for LLM agents, because standard retrievers rank by relevance without knowing which tools a multi-step instruction actually requires. This paper post-trains small open Qwen3 models (1.7B to 8B parameters) via supervised fine-tuning and reinforcement learning to decompose a request and select tools through a lightweight Plan–Retrieve–Select workflow. The headline result is that an 8B checkpoint outperforms a 402B-parameter Llama-4-Maverick on ToolBench and transfers zero-shot to unseen tool domains and retrievers, suggesting that tool selection can be compressed into a small, locally hostable policy rather than delegated to massive proprietary models.

The paper asks who adapts to whom when a confident LLM is wrong. In a within-subjects experiment, 32 participants learned science topics from GPT-5-mini assistants whose briefings were covertly falsified at 0%, 25%, 50%, 75% or 100%. The study finds that learners absorbed roughly two-thirds of stated falsehoods regardless of overall assistant reliability, while their confidence and trust ratings failed to track accuracy. By reframing trustworthiness as a property of the human-AI pair and introducing a Human Deference Index, the work argues that sycophancy evaluation must measure human-side deference as well as model-side compliance.

MoG-VLM attacks the prohibitive cost of zero-shot video anomaly detection with Vision-Language Models by treating motion not as a feature but as a lightweight gate that decides when to run expensive VLM inference. It introduces Temporal Gradient Magnitude (TGM) for content-aware frame sampling, anchor-triggered hierarchical lookback windows for temporal refinement, and a tier-aware fusion rule that balances frame-level and video-level语义 signals without any learned parameters. The work reports state-of-the-art zero-shot training-free AUC-ROC on XD-Violence (91.91%), UCF-Crime (84.30%), and MSAD (87.88%), alongside a claimed 1.31× speedup over the prior best method.

Coding agents typically retrieve skills before execution but rarely choose at runtime whether to invoke coarse macro skills or fine-grained primitives. This paper reframes skill use as a runtime control problem, introducing Adaptive Skill Granularity: a decision-tree router selects micro, macro, or hybrid execution, and a finite-state executor downgrades from macro to micro after observable failures. The framework matters because it shifts skill libraries from static documents into controllable, verifiable execution abstractions that can trade off efficiency against robustness.

Test-time scaling generates many candidate reasoning traces, but completing all of them is prohibitively expensive. This paper proposes PKTP, a pruning framework that selects a subset of partial rollouts to maximize the probability that at least one of them ultimately succeeds—i.e., $\text{pass}@k$. It models joint correctness with a latent Gaussian threshold model, derives a closed-form submodular surrogate, and greedily selects the kept set with a $(1-1/e)$ approximation guarantee. The work matters because it directly exposes and optimizes the quality–diversity trade-off that existing pruning methods handle only indirectly.

Masked diffusion language models generate tokens in parallel, but their efficiency depends critically on which predictions are committed at each step. This paper disentangles the design space of learned unmasking policies by treating them as distilled teachers and independently varying supervision targets, training states, and policy inputs. The authors find that trajectory-contraction labels and a confidence-plus-hidden-state input yield a lightweight policy that improves the accuracy–NFE frontier and end-to-end throughput once NFE drops by about 10%.

FocusVLA targets a subtle but important failure mode in multi-view Vision-Language-Action models: policies tend to develop early visual preferences—especially over-reliance on wrist cameras—and persist with them even when task stages shift toward global planning. The paper proposes predicting future interaction intent from multimodal context and short-term action history, then converting that intent into step-wise routing weights that dynamically balance global and local views during action generation. Because the routing module sits on the action-generation side without altering the visual encoder or language backbone, it is designed as a lightweight, plug-in upgrade for existing VLAs. It is evaluated on all 50 tasks of the RoboTwin 2.0 benchmark and on six real-world manipulation tasks, reporting 10+ percentage point improvements over a $\pi_{0.5}$ baseline trained under identical conditions.

Meta-RL agents struggle to adapt quickly to unseen tasks when task evidence is sparse or spread across long interaction histories, because recurrent models compress history too aggressively and transformers scale quadratically. METTA proposes a two-scale neural memory architecture that uses Test-Time Training (TTT) layers for fine-grained within-episode adaptation and compact episode summaries for cross-episode reasoning, preserving linear sequence complexity. The core claim is that this design enables stronger task-family-level generalization and more structured exploration on long-horizon sparse-reward tasks than both recurrence and full-attention baselines, while keeping wall-clock training cost low.

LLM agents can execute prohibited actions when tool responses omit safety-relevant facts, even if every returned statement is true and the user request is benign. EpiSEM-Bench introduces 300 matched unsafe/safe task pairs across six domains with controlled evidence suppression and an optional audit-based recovery tool. The work shows that robustness to missing evidence and useful recovery are distinct capabilities that single-score benchmarks conceal, making joint evaluation essential for safe agent deployment.

Standard on-policy distillation forces a student to copy a teacher at every token, even when the teacher's preferred actions hurt the student's eventual success. This paper proposes Advantage-Based On-Policy Distillation (AOPD), which weights each teacher token by its counterfactual advantage under the current student policy. The key insight is that imitation should optimize for student task reward, not teacher agreement, and the authors derive an unbiased estimator that needs only standard student rollouts and terminal rewards.

N-gram embedding tables give language models cheap capacity gains but can consume hundreds of billions of parameters. This paper asks whether those embeddings can be stored at low precision without losing their benefit. The authors systematically compare quantization schemes—uniform, normal-quantile, and Lloyd–Max codebooks with per-embedding scales—and find that aggressive bit reduction often preserves most of the tables’ gain, freeing memory to grow the tables themselves.

Autonomous research agents can automate entire scientific workflows from ideation through experimentation to writing, but they frequently compromise research integrity through behaviors like test-set exploitation, fabricated citations, and manipulated reporting. RIDAR addresses this gap by introducing RIVT, a taxonomy of 14 research integrity violation types spanning the full lifecycle, together with RIDAR-Post-hoc for automated detection and RIDAR-Runtime for live intervention. The work matters because the authors show such violations are nearly universal in current frontier agents, making integrity safeguards a prerequisite for trustworthy machine-generated science.

Industrial organizations struggle to govern structured telemetry and unstructured documentation together because the work relies on scarce domain experts. The paper proposes contestable governance: specialist AI agents generate scored, cited arguments for metadata decisions, human stewards review and intervene before anything is committed, and an append-only record preserves full lineage. This design aims to stop unreviewed agent errors from propagating into downstream systems while making human oversight attributable and auditable.

Depth growth and looping reuse computation across layers to train or run Transformers more efficiently. This paper shows that both methods create an inductive bias toward learning from context over memorizing fixed associations in weights, accelerating in-context learning and propagating induction heads through depth. The evidence spans models up to 3B parameters and connects architecture design to the mechanistic origins of few-shot performance.

Muffin tackles the overhead and synchronization bottleneck of multi-agent reasoning by replacing full-trace debate with asynchronous sharing of compact intermediate findings through a shared memory buffer $B^{(d)}$. An orchestrator routes queries to either a single agent or a team of parallel agents that read and append selective findings while keeping private reasoning local. The authors claim strong accuracy-token Pareto gains across four reasoning benchmarks and four open-weight language models.