Policy-Verified Self-Training asks whether a runtime policy gate's discarded verdicts can be recycled as training signals for the same agent. The paper proposes PVST: an agent rolls out trajectories, a deterministic gate compiled from written policy judges compliance, and the verdicts become acceptance filters and preference labels. On AgentDojo prompt-injection benchmarks, post-training Qwen3-8B with these verifier-derived rewards cuts violating trajectories from 37.2% to 10.3% and successful attacks to 0.5% without any stronger teacher model, human preference labels, or attack-specific supervision. The result suggests that compiled policy checkers can serve not only as runtime constraints but as a scalable, auditable source of self-training supervision.

LLM agents frequently fail in multi-turn settings by violating policies, using tools incorrectly, or acting prematurely. This paper proposes Critique-Augmented Agentic Fine-Tuning (CAFT), which rolls out a student agent, has an expert label each wrong turn with both a textual critique and a corrected action, and then fine-tunes the student to generate that critique and correction while keeping the original mistake in context. The central finding is that these natural-language explanations provide learning signal beyond action-only supervision, improving task success across four models and two benchmarks.

Multimodal agents need persistent memory to act coherently over long horizons, yet forming and managing memory across heterogeneous, high-dimensional observations—video, audio, 3D geometry, and GUIs—has lacked a unifying analytical lens. This survey introduces a two-axis lifecycle framework that classifies how experience is represented (source, textual, structured, latent, or parametric) and how memory states and policies evolve ($T1$ Static, $T2$ Dynamic, or $T3$ Self-Evolving). By analyzing 297 architectures, the authors identify key trade-offs between fidelity and efficiency and expose the striking scarcity of systems that evolve their own memory policies at deployment time.

DangerLens asks whether two language models with identically safe refusal behavior can still differ in how much hazardous knowledge remains internally accessible. The authors train paired Qwen1.5 checkpoints—one directly safety-aligned, the other first exposed to harmful data then aligned to the same standard—and introduce a vocabulary-space probe that uses a Jacobian lens and blinded semantic scoring to count harmful goals and operational means inside hidden states. The work matters because current safety audits rely on surface behavior alone, leaving latent dangerous representations undetected before downstream fine-tuning or attacks.

Best-of-N sampling improves LLM output quality by generating multiple candidates and selecting the highest reward, yet its scaling behavior under imperfect reward models is poorly understood. This paper recasts BoN as extreme-value competition between the upper tails of correct and incorrect reward distributions, yielding phase transitions that explain both monotonic gains and reward hacking. Using extreme value theory, the authors derive finite-sample scaling laws showing polynomial rather than exponential convergence, and propose a lightweight tail-tension criterion for label-free early stopping. The work matters because it replaces vague intuitions about reward hacking with precise, testable distributional conditions and a practical inference-time safeguard.

Activation steering is used to steer language models toward named behaviors, but standard baselines cannot tell whether a vector controls a specific concept or a neighboring one. This paper introduces the competing-target test, which compares a steering vector with the matched vector for a competing target and bounds their difference with an equivalence test. In three open-weight models, the authors show that vectors for Kantian deontology and act utilitarianism are largely interchangeable---they move both framework scores alike---whereas explicit prompting separates them, and the same construction does separate optimistic and pessimistic tones.

This paper studies the infinite-generation limit of autoregressive transformers. Rather than fixing sequence length, the authors let generation time tend to infinity and ask whether the model's empirical state converges. Their core result is that, for a broad class of architectures, the stochastic generation process tracks a deterministic ODE on a slow time scale, and every fixed transformer has a finite temperature threshold above which it converges to a unique limit. However, no uniform temperature works for all models; for any finite temperature, some transformer can fail to converge, even with just one layer and three tokens.

AI agents chain many LLM calls to complete tasks, yet models are usually compared per call. This paper argues the correct functional unit is the successfully completed task, because failed attempts waste energy, cost and time. It derives crossover task lengths that predict when a smaller model stops being greener, cheaper or faster, and shows with a 15-model pilot that the greenest choice shifts from 1-3B models for short tasks to 8-27B models for long ones.

Agents are increasingly expected to negotiate and collaborate across organizational boundaries where each party maintains private tools, documents, and incentives. This paper studies how agent-to-agent (A2A) systems fail under these conditions, introducing A2A²-Failures, a taxonomy of 20 failure modes derived from over 2,500 sandboxed runs spanning cooperative to competitive scenarios. Its central thesis is that evaluating only final outcomes masks critical behavioral breakdowns—such as unverified peer claims, private data leakage, and scope distortion—making per-agent principal-alignment auditing essential for safe cross-boundary deployment.

LLM-jp-4 is a family of Japanese sovereign language models—an 8B dense Transformer and a 32B-A3B mixture-of-experts—trained from scratch on roughly 12 trillion tokens with an explicit emphasis on transparency, legal reproducibility, and practical deployment. The authors impose strict open-source selection criteria on data, external models, and software, and they document every stage from tokenizer design through pre-training, mid-training, and post-training. The work aims to provide a publicly verifiable reference for sovereign AI development in Japan.

Minimax optimization underlies adversarial training, GANs, and AUC maximization, yet biased stochastic gradient descent-ascent methods—zeroth-order and clipped variants—have had no generalization theory. This paper develops the first stability framework for biased SGDA in convex-concave smooth problems, replacing the standard bounded-gradient assumption with a bounded second-moment condition. It derives generalized Lipschitz-type stability bounds and proves that zeroth-order and clipped SGDA can achieve the minimax-optimal $O(1/\sqrt{n})$ excess risk.

Real-time monocular human pose estimation breaks down when the lower body is occluded or the subject steps out of frame. This paper introduces a part-decoupled temporal fusion framework that augments monocular video with sparse plantar-pressure and foot-mounted acceleration from smart insoles. The core insight is that cheap, wearable foot sensors provide complementary physical constraints exactly where monocular vision is weakest, enabling robust pose and gait recovery under severe visual degradation.

FrameEvidence is a diagnostic benchmark for multi-image spatial reasoning that tests whether vision-language models correctly answer questions when the reference frame or the set of visual evidence is manipulated. It introduces 960 Chinese multiple-choice questions arranged in 320 matched triplets, where FQ1 varies the spatial reference frame while holding images fixed and FQ2 deletes or adds images to test evidence necessity and invariance. The core finding is a sharp divergence between single-item accuracy, which reaches 50\% for GPT-5.6, and joint correctness across matched variants, which remains below 11\% for $\mathrm{FQ1\text{-}GEM}$. This design exposes how standard accuracy obscures brittle reasoning and answer instability.

This paper tackles the problem of recursive self-improvement for agentic video-understanding harnesses, asking how automated evolution can discover generalizable designs when it must learn from limited, noisy, and repeatedly reused validation feedback. The authors first diagnose three interconnected failure modes in a baseline system (validation overfitting, search-budget dilution, and the early-experience trap), then propose VideoHarness-SRSI, an evidence-governed framework that hides validation content from the proposer, warms up with controlled interventions, and evolves mechanisms in isolated probe–race branches before factorial composition. Across four benchmarks, the method improves the mean held-out score by 6.2% relative to the baseline and cuts the validation–test gap by 37.9%, supporting the central claim that how search is informed and organized matters as much as how well it optimizes validation metrics.

Wedge addresses the inefficiency of generic environment generation for embodied navigation: pretrained agents fail for different environmental reasons, so simply adding more diverse or realistic training scenes may not address their actual shortcomings. The paper proposes a policy-conditioned generation framework that diagnoses an agent's specific weaknesses by contrasting its successful and failed rollouts, uses an LLM to propose executable environmental features (such as target visibility or blind-path ratio), validates them via success-rate gaps, and generates targeted training environments. This approach is tested across three heterogeneous agents—an RL policy, an imitation-learned model, and a black-box VLM—showing consistent gains in ProcTHOR and transfer to HM3D.

When text and images conflict in vision-language models, prior work has characterized behavior as modality-specific biases such as "blind faith in text." This paper proposes a predictive alternative: a two-parameter Bayesian cue combination rule in which each modality contributes conditionally independent evidence weighted by a constant integration bias, plus a prior. The authors show that this simple linear model accurately predicts how VLM preferences shift across dozens of input degradation conditions and even transfers to entirely new corruption types without refitting, offering a unified account of why text sometimes dominates and vision sometimes wins.

Equivariant GNNs for molecular modeling are bottlenecked by Clebsch-Gordan tensor products, which suffer from high complexity and limited non-linearity. This paper proposes the Equivariant Spherical Transformer (EST), a plug-and-play module that applies self-attention over sampled points on the spherical spatial domain of group representations, using a Fibonacci Lattice to preserve approximate SO(3) equivariance. The work matters because it reports state-of-the-art results on OC20 and QM9 while empirically breaking the degree-limited expressiveness barrier that constrains traditional tensor-product architectures.

Most LLM agents treat repeated use of the same external source as a sequence of independent retrieval events. This paper reframes the deeper problem as source learning: progressively building a persistent, reusable source model that captures how a source’s knowledge is structured, interpreted, and applied. The proposed SourceLearn system combines self-directed study with task-guided refinement, and the authors report large gains across document, code, API, and interactive benchmarks.

The paper tackles how to teach LLMs new skills after pretraining without erasing old ones. While RL resists catastrophic forgetting by staying on-policy, it struggles when the model cannot generate successful trajectories on its own. The authors flip the problem: instead of changing the training loss, they use an MCMC projection sampler to reshape off-policy expert traces so they look more like samples from the base model while preserving correctness. This lets ordinary supervised finetuning match or beat RL and self-distillation on science, math, and medical tasks.

FedMIX-P tackles preconditioner coordination in federated learning by mixing a shared reference preconditioner with each client's local adaptive state at every local step. The core idea is to control the operator mismatch that arises when heterogeneous clients transform gradients differently, and the paper demonstrates empirical improvements on vision and language models under severe data heterogeneity. However, the convergence analysis is restricted to linear positive-definite preconditioners, while the headline experiments rely on nonlinear methods such as Muon, creating a notable gap between theory and practice.