3D object composition in simulated environments demands more than static coordinate prediction; complex tasks like hanging, containment, and soft-body settling require iterative interaction and dynamic physical feedback. PhySGA addresses this by embedding LLM/VLM agents directly inside a physics simulator, where an Assembly Agent plans coarse metric adjustments and a Critic Agent refines them via pixel-grounded visual feedback, supported by modular Action Routers and a Self-Evolving Skill Library. The framework matters because it moves beyond layout priors and treats physical construction as a closed-loop agentic process applicable to rigid, articulated, and deformable objects.

SwarmRSI tackles recursive self-improvement in scientific research by treating hypotheses, programs, and experimental evidence as inheritable states that evolve across a swarm of LLM agents. Its core idea is a shared research graph enabling feedback-guided mutation, recombination of partial solutions across agents, and adaptive allocation between hypothesis and program refinement. The framework reports state-of-the-art scores on OpenProblems and NatureBench and provides ablation and trajectory evidence that cross-agent reuse drives progress.

Video large language models often hallucinate answers that are not supported by the input footage. Contrastive decoding can reduce this by comparing the original video against a weakened "negative" view, but existing negatives are either predefined or optimized for query reconstruction rather than the model's actual visual reliance during answer generation. ARP-CD introduces a training-free fix: it treats the attention mass on visual tokens at the answer-prediction position as a differentiable optimization target, then uses bounded gradient ascent to perturb the input visual hidden states and create a negative branch for contrastive decoding. The authors evaluate the idea across four Video LLMs and four benchmarks, supplementing headline results with ablations and a mechanistic analysis of first-token logit margins.

Cooperative success depends on implicit conventions, yet conventions are usually inferred only indirectly from score drops when partners are swapped. This paper introduces the convention gap $\Delta = \operatorname{mean}(P(\text{life loss})) - \text{actual life-loss rate}$ in the game Hanabi, using it to identify which situations carry implicit meaning for humans and for a hierarchy of off-belief learning agents, and to measure what removing those conventions costs. By comparing populations on a single exact literal baseline, the authors turn conventions from inferred causes of failure into directly measurable and causally testable objects.

On-policy distillation (OPD) trains a student on its own rollouts using token-level supervision from a teacher, but its effectiveness relies on the teacher's ability to recover from any student prefix. The authors show that this recoverability collapses on incorrect rollouts: the teacher follows the student's errors rather than correcting them, even though it can still recognize them. Their method, CriticOPD, restores recoverability by having the teacher critique an incorrect rollout so the student can rewrite its solution from the last error; the teacher scores the rewrite with the critique as privileged information, while the student trains on the revision without it. On six mathematical reasoning benchmarks, CriticOPD improves standard OPD by roughly 4–6 points across two student scales, with the largest gains on competition-level problems, and the student needs no feedback at inference.

Coding agents outgrow their context windows and compact older history into summaries, silently dropping commitments and facts that lightweight models cannot recover. This paper introduces transparent-compact, a deterministic compaction method that keeps all user and assistant prose verbatim, explicitly flags removed tool outputs and reasoning, and delegates recovery in a single lookup step. The core claim is that sparing lightweight models from guessing around lossy summaries improves retention of rules and facts without task-specific engineering. The evidence is strongest for local models like Qwen3.8-27B and weaker for frontier models, with a clear ceiling on how many verbatim rounds are possible before truncation.

PerSo introduces a benchmark for evaluating LLM social intelligence through real-world persuasion dialogues from Reddit's ChangeMyView. It tests cognitive intelligence by forecasting a persuadee's beliefs, emotions, and intents, and behavioral intelligence by selecting the more persuasive response. Unlike prior work that often relies on synthetic scenarios or isolates these two aspects, PerSo jointly evaluates both on naturally occurring interactions, providing a more ecologically valid testbed.

This paper identifies a gap between standard visual grounding—which assumes enough dialogue context to pinpoint a single object—and real-world interactions where multiple referents remain plausible. It proposes Candidate Set Visual Grounding (CSVG) and introduces the CandVG benchmark to evaluate how the set of plausible referents evolves across dialogue turns. Evaluating 18 vision-language models shows that stronger single-object grounding does not translate to better set estimation, with models often failing to generate full candidate sets or committing prematurely to one candidate.

Underwater vessel recognition is limited by scarce training data and small public benchmarks. This paper introduces ONCship, a 4521-hour corpus curated from Ocean Networks Canada hydrophones and Automatic Identification System vessel metadata. It evaluates domain-specific self-supervised pre-training, supplementary supervised fine-tuning, and MMSI-Level individual vessel classification, showing that scale and metadata can improve representations even though downstream gains are architecture-dependent and uneven.

Potato leaf diseases such as early and late blight reduce yields in Ethiopia, where farmers often lack access to laboratory diagnosis. This study develops a lightweight convolutional neural network to classify healthy, early-blight, and late-blight leaves, training on 2,215 PlantVillage images and testing on 247 smartphone photos collected from local farms. The central claim is that a 1.70 MB model can achieve high accuracy while remaining efficient enough for deployment on low-resource hardware.

Most robot VLAs rely on wrist cameras for manipulation detail, which narrows scene coverage and adds hardware complexity. FoveaVLA removes that dependency by teaching a π0.5-based policy to predict future interaction regions, cropping high-resolution 'foveal' patches from external stereo views to recover fine gripper-and-object detail. Post-trained on DROID with RoboInter annotations, the model achieves 77.1% overall real-world success with two external RGB-D views versus 52.1% for the wrist-equipped baseline. The idea is biologically grounded and architecturally clean, though the evaluation scale and claims about learning from human videos remain tentative.

Fairness-weighted federated learning deliberately inflates updates from high-loss clients, but under an update-only interface a server cannot tell legitimate amplification from adversarial scaling. This paper identifies that $q$-fair loss scaling preserves the class-mixture structure of honest updates while warping their coefficient constraint from a unit-sum simplex to a radial nonnegative cone. Building on this geometry, FUSE uses a small trusted labeled set to recover latent coefficients, screens submissions jointly for radial scale and cone consistency, and renormalizes the aggregate to protect both robustness and fairness.

Most research on test-time compute treats the model as fixed and asks how to allocate extra inference tokens. This paper instead asks whether a model can be post-trained to use a given compute budget more effectively. The authors identify compute-responsive (CR) problems—those that flip from complete failure to success under additional blind re-solving or audit-based revision—and show that supervised fine-tuning on these examples shifts the compute–accuracy frontier across budgets from 512 to 4,096 tokens and generalizes across model families.

Training deep networks without normalization remains difficult, and relating their trainability to the exact geometry of the loss surface is rarely attempted. This paper proposes a projective polynomial activation $f(z)=z+\lambda z^3$ paired with either an $\ell_2$ projection onto the unit sphere or an unnormalized identity path, and studies it at two scales: networks that train to depth $100$--$300$ on MNIST, and tiny algebraic versions whose square-loss critical points are solved exactly via Gröbner bases and numerical irreducible decomposition. The point is to derive concrete design, initialization, and diagnostic rules from the small model's geometry and validate them at depth.

Multimodal code intelligence demands programs that are correct, well-structured, and visually faithful, yet existing training signals are either too coarse or confined to single tasks. This paper introduces PrismaForge, which synthesizes per-instruction evaluation rubrics that separately assess code-level and visual-level criteria, then uses them to filter training data and reward reinforcement learning. The resulting dataset, PrismaCoder-2M, and the trained 9B and 35B models advance the open state of the art and compete with proprietary systems on charts, webpages, and scientific graphics.

Deploying LLMs as agents requires respecting runtime-specific execution contracts, yet standard training leaves harness semantics implicit in action trajectories. This paper introduces Harness Policy Distillation (HPD), which replays reference trajectories through the target harness to compile control-state documents, uses harness verification to filter teacher actions, and jointly trains a shared model on native action prediction, document-conditioned action prediction, and document reconstruction. On TRAJECT-Bench with the Pi harness, HPD lifts Qwen3-8B task success from 3.09% to 22.98% and substantially outperforms supervised, RL, and standard distillation baselines.

Average-reward reinforcement learning with options models continuing hierarchical tasks, but before this paper there were no finite-time convergence guarantees for option-based Q-learning in that setting. The authors prove that multi-step (J-step) inter- and intra-option Bellman operators contract in the span seminorm, design corrected synchronous multi-step estimators, and derive the first $\widetilde{\mathcal{O}}(1/k)$ mean-square convergence rates for average-reward option Q-learning. The results extend to hierarchical reward machines and are supported by small-scale tabular experiments.

Speech Emotion Recognition (SER) models and Audio LLMs often suppress vocal emotion when it contradicts transcribed text—a bias the authors label Semantic Dominance. To probe this, they introduce ASPIRE, a fully synthetic benchmark of adversarial audio-text pairs generated via EmotiVoice, and propose ACR-Net, a dual-stream Whisper-based architecture that uses a Contrastive Decoupling Loss to push conflicting acoustic and semantic representations toward orthogonality. The work claims state-of-the-art Acoustic Accuracy on both adversarial and standard congruent datasets.

Consensus research has expanded from human social networks to multi-agent LLM systems, yet it remains unclear how language models coordinate when their preferences conflict. This paper adapts the Network Consensus game—where agents must agree on one of two colours despite receiving different payoffs for the agreed colour—to evaluate seven LLMs under varying network topologies and communication regimes. The authors find that open-source models reach consensus far less often than humans and usually require global messaging to do so, whereas ChatGPT 5.6-sol succeeds even with restricted communication through lower stubbornness and stronger majority-following. The work matters because it directly challenges the assumption that LLMs can serve as drop-in proxies for humans in social simulations or collective decision-making architectures.

This paper investigates how Retrieval-Augmented Generation affects question answering across five South Asian languages. The authors construct a Wikipedia-based multilingual dataset, generate questions with Qwen 3 32B, and compare non-contexted, contexted, and RAG-based answering using four LLMs. They find that RAG improves QA unevenly—substantially aiding non-popular cultural questions in Bangla and Tamil while underperforming direct contexted answering in Hindi and Nepali—suggesting that retrieval quality and language-specific data alignment matter more than the choice of RAG framework.