Zeolite synthesis requires a structured recipe that specifies an organic structure-directing agent (OSDA), gel composition, crystallization temperature, and reaction time. General-purpose language models can draft such recipes, but their outputs often fail domain checks, producing invalid SMILES strings, missing condition fields, or temperature–time schedules inconsistent with local Arrhenius trends. We present ZeoAgent, a closed-loop compiler for machine-checkable zeolite synthesis recipes. ZeoAgent represents each generated recipe as a typed record and verifies it through deterministic field-level checks. At each iteration, a compact yet diverse set of retrieved synthesis precedents guides the next draft. When a field fails validation, the corresponding error code is returned to the model for targeted revision. Recipes passing these checks are calibrated with a local Arrhenius fit before final ranking. In condition-only experiments, the recorded OSDA is supplied and only the hydrothermal-condition fields are predicted. On the archived model outputs, ZeoAgent attains 99.7% RDKit parseability, compared with 81.7% for Qwen3-8B. Typed feedback raises repair pass@5 from 68.3% to 72.0%, and PGC reduces high-residual schedules under its fitted local Arrhenius self-consistency diagnostic from 77.0% to 5.4%.

Self-distillation in reinforcement learning with verifiable rewards (RLVR) uses successful in-group rollouts to teach the failed ones, with no external teacher needed. Whether it actually helps turns on three closely related design questions that the canonical method does not handle carefully: how to construct an appropriate teacher context for a given loser, where in the response the distillation loss should act, and what kind of signal it should carry. The canonical Self-Distillation Policy Optimization (SDPO) instantiation answers all three by default: a random in-group winner, a global distillation loss on the whole response, and a Jensen–Shannon divergence on the full top-K distribution. These three defaults compound into a catastrophic late-training failure that short-horizon evaluation hides. We propose CPSD (Counterfactual Patches Self-Distillation), which replaces each default in turn: pick the teacher by minimum embedding-weighted edit distance, localise distillation to an adaptive-stop counterfactual patch beginning at the first divergent token, and replace JS with a Plackett–Luce listwise loss that preserves only the teacher’s ranking. Across six reasoning benchmarks at a fixed 150-step budget, CPSD eliminates the SDPO cliff and lifts average accuracy by +5.7 pp over GRPO on Qwen3-8B and +4.2 pp on Qwen3-4B.

This paper explores agentic 3D spatial understanding, i.e., MLLM agents performing indoor 3D reasoning through tool use. Existing methods often misuse tools and exhibit biased tool preferences, leaving the agentic paradigm with only marginal gains over non-agentic strategies. We reveal that 3D spatial reasoning tasks are heterogeneous across scenes, while these agents apply a uniform tool-use strategy to all scenes rather than selecting tools according to the specific scene and task. To address this, we propose Skill-3D, a framework that learns self-evolving scene-aware skills. Specifically, Skill-3D identifies the task scene and records the agent’s tool-use trajectory into a Scene Memory, where successful trajectories from similar scenes are aggregated and distilled into a reusable scene-aware skill, with failed ones attached to the skill as lessons. During training, once a similar scene recurs, the corresponding skill is injected to guide the agent, producing new trajectories whose successes and failures further refine the skill, forming a loop in which the memory and the skill library co-evolve. Experiments show that Skill-3D substantially improves tool utilization in 3D spatial reasoning (from 39%to 78% on VSI-Bench), driving the agent toward correct and sufficient tool use. For instance, it improves Gemini-3-Flash by 67% on MMSI-Bench. Furthermore, we conduct agentic reinforcement learning over skill-guided trajectories, which boosts Qwen3-VL-8B by 60% on VSI-Bench.

Reinforcement Learning (RL) has proven effective in eliciting reasoning abilities of Multimodal Large Language Models (MLLMs) and is emerging in document understanding. However, existing approaches usually lack fine-grained region-level grounding supervision, while the key challenge in document reasoning lies in linking visual evidence within precise regions directly to question–answer semantics. We introduce RegionDoc-R1, a novel framework for document understanding that enhances MLLMs’ reasoning with step-wise feedback. Instead of previous OCR-based text positions, we encourage the model to predict question-relevant Regions of Interest (RoIs) and ground the answer in these regions. Specifically, we first propose the Region-Aware Group Relative Policy Optimization (RA-GRPO), which encourages models to utilize region-level spatial information in documents for reasoning. Correspondingly, we construct a hybrid training corpus, named SR-Doc, containing Semantic Reasoning (SR) examples enriched with cross-page and region-level reasoning layout annotations. Meanwhile, we also introduce an Adaptive Chain-of-Thought (Ada-CoT) strategy, which dynamically adjusts the reasoning process according to different tasks, enabling more efficient and flexible step-wise document understanding. Experiments on several document reasoning benchmarks demonstrate that RegionDoc-R1 achieves state-of-the-art performance across tasks such as form document understanding, table-based QA, and layout-sensitive information extraction.

Large language model (LLM)-based multi-agent systems (MAS) improve reasoning through specialized collaboration, but dense multi-round communication can increase prompt cost and expose agents to outdated, redundant, or noisy context. A prominent line of MAS communication-efficiency work reduces communication through agent-, edge-, or topology-level structural decisions, which operate on graph structure rather than the realized text entering a specific receiver prompt. Such structural decisions are too coarse for runtime communication: messages sent over the same edge may carry useful evidence in one query, redundant content in another, or only a few useful spans. We introduce MessageSieve, a training-free runtime message pruning policy for fixed-graph MAS that operates on message instances rather than agents or edges. Before prompt assembly, it uses lightweight state cards and local message comparisons to preserve original text, retain compact evidence, or omit messages, without a trained selector or additional LLM call. We evaluate MessageSieve on two backbone models and five benchmarks covering expert QA, math reasoning, and code generation. Experiments show that MessageSieve improves the macro-average score over dense Full MAS by 4.10 points on Qwen3-8B and 2.59 points on Gemini 3.1 Flash-Lite, while reducing evaluation tokens by 19–49%.

Large Language Models (LLMs) often fail to utilize their latent reasoning capabilities due to a distributional mismatch between ambiguous human inquiries and the structured logic required for machine activation. Existing alignment methods either incur prohibitive O(N) costs by fine-tuning each model individually or rely on static prompts that fail to resolve query-level structural complexity. In this paper, we propose ReQueR (Reinforcement Query Refinement), a modular framework that treats reasoning elicitation as an inference-time alignment task. We train a specialized Refiner policy via Reinforcement Learning to rewrite raw queries into explicit logical decompositions, treating frozen LLMs as the environment. Rooted in the classical Zone of Proximal Development from educational psychology, we introduce the Adaptive Solver Hierarchy, a curriculum mechanism that stabilizes training by dynamically aligning environmental difficulty with the Refiner’s evolving competence. ReQueR yields consistent absolute gains of 1.3%–7.2% across diverse architectures and benchmarks, outperforming strong baselines by 2.1% on average. Crucially, it provides a promising paradigm for one-to-many inference-time reasoning elicitation, enabling a single Refiner trained on a small set of models to effectively unlock reasoning in diverse unseen Solvers. Code is available at https://github.com/newera-xiao/ReQueR.

Sparse Mixture-of-Experts (SMoE) architectures are widely used in large language models (LLMs) due to their computational efficiency. However, though only a few experts are activated for each token, SMoE still requires loading all expert parameters, leading to high memory usage and challenges in deployment. Previous work has tried to reduce the overhead by pruning and merging experts, but primarily focused on expert-level operations, leaving neuron-level structure underexplored. We propose DERN (Dropping Experts, Recombining Neurons), a task-agnostic and retraining-free framework for expert pruning and reconstruction. We observe that experts are often misaligned and contain semantic conflicts at the neuron level, which poses challenges for direct merging. To solve this, DERN works in three steps: it first prunes redundant experts using router statistics; then it decomposes them into neuron-level expert segments, assigning each segment to its most compatible retained expert; and finally, it merges segments within each retained expert to build a compact representation. Experiments on Mixtral, Qwen, and DeepSeek SMoE models show that DERN improves performance by more than 5% on commonsense reasoning and MMLU benchmarks under 50% expert sparsity, without extra training. It also greatly reduces the number of experts and memory usage, making SMoE LLMs easier to deploy in practice.