Agent scaffolds, circuit concentration, and reasoning coverage floors

Nguyen Xuan-Vu, Octavian Susanu, Daniel Armstrong, Philippe Schwaller

MAELLE models chemical reactions as discrete flow matching over electron occupation vectors on bonding, non-bonding, and hydrogen sites rather than product generation or heuristic graph edits. It constructs intermediate trajectories via optimal transport on a continuous-time Markov chain, recovering interpretable mechanistic edit sequences without elementary-step labels. The method matches leading USPTO-480K accuracy, stays robust under structural and reaction-type shifts, and naturally surfaces side products from full electron redistribution.

Sil Hamilton, Albert Yu Sun, Oscar J. Romero, Carl-Leander Henneking, David Mimno, Bishan Yang, Igor Labutov

CorporateBench constructs human-validated multi-task Q&A corpora that reach realistic enterprise scale, exceeding 230,000 documents drawn from temporally evolving, logically consistent synthetic firms of 12 to 10,000 employees. It evaluates information extraction and knowledge-base querying under conditions that mirror corporate communication networks. Five LLMs show sharply degrading performance as corpus size approaches real-world scale, exposing a critical gap in existing benchmarks.

Allison Zhuang, Santiago Aranguri

The paper decomposes verbalized eval-awareness in chain-of-thought into capabilities-flavored, safety-flavored, both, or neither framings and shows they predict compliance very differently. On Qwen3-32B and FORTRESS, capabilities framing yields a 24 to 46 percentage-point compliance gap over safety framing across steering conditions. CoT-prefill interventions confirm causality, demonstrating that aggregate suppression of eval-awareness can leave the safety-relevant component untouched.

Yuan Chang, Xiaoqi Chen

NPO is a lightweight single-lineage method that iteratively revises prompts with a teacher model and rollout feedback, matching or beating GEPA with fewer rollouts. Gains increase with stronger teachers, indicating that teacher reasoning can substitute for optimizer search complexity. Optimized prompts transfer across student models within the same family, showing that simple linear optimization can rival elaborate multi-lineage search for recursive self-improvement.

Xiaokun Guo, Zhen Xu, Dongdong Huo, Yanqiu Zhang, Wei Wang, Qinfu Yang, Dongjin Yu, Yu Wang

SARA treats action induction and execution authorization as distinct runtime roles, using a context-isolated Action Probe to record provenance while authorizing tool calls only against user objectives and audited evidence. No-History-Promotion prevents historical recurrence from converting induced actions into authorized ones. Across AgentDojo and AgentDyn the method caps attack success rate at 0.63 percent while preserving competitive task utility.

Zike Yuan, Han Zhang, Jianzhi Yan, Le Liu, Cai Ke, Huozhi Zhou, Jian Xie, Jiran Yin, Yukun Cao, Yue Yu, Hui Wang, Ming Liu, Bing Qin

GRAIN is a single-agent RL framework that casts graph reasoning as semantic parsing plus tool execution and trains with a Structure Invariance Reward that validates intermediate graphs against ground-truth topology. This forces robust text-to-structure mappings instead of surface-pattern memorization. It outperforms multi-agent baselines by 16.45 percent accuracy at lower latency and halves the OOD gap of SFT models on the new GRIT benchmark of linguistic shifts.

Xinwei Qiang, Xiang Fang, Chang Chen, Yue Guan, Yufei Ding

The authors separate rejection in block drafters into an information floor (minimum expected rejection given conditioning order) and a residual model gap. Estimating both from target rollouts shows the all-parallel floor reaches 0.286 at the final slot, while one realized token removes 86 to 100 percent of that floor. Current drafters sit far above their floors, with model gap accounting for most of observed rejection, clarifying the value of short-range conditioning versus proposal quality.

Sai Adith Senthil Kumar

Circuit Condensation iteratively prunes low-attribution edges and trains low-rank adapters to restore performance through the remaining graph, retaining cuts only when task and general capability survive. Across four behaviors and eight models it shrinks circuits 8.1 times on average (up to 316 times) relative to frozen discovery baselines, with weight updates driving the reduction. Exhaustive subset tests and pair ablations confirm many edges are removable and reveal inter-edge dependencies, yielding cleaner sufficient sub-circuits that track original next-token distributions.

Xingyu Shen, Huishuai Zhang, Peng Li, Yinchun Wang, Dongyan Zhao

The method forces a strong policy to continue from partial reasoning trajectories produced by a weaker model, injecting distributional discrepancy that disrupts over-confidence and expands reasoning coverage during RLVR. This outer-prefix guidance preserves generative diversity without extra SFT, complex rewards, or intricate prompting. Gains grow with larger k on mathematical benchmarks, showing substantial expansion of pass@k coverage while mitigating entropy collapse.

Siye Wu, Kai Yang, Yuchen Cai, Xin Xu, Peng-Yuan Wang, Jiaxuan Wang, Jiashun Liu, Jiafei Lyu, Yangkun Chen, Saiyong Yang, Yanghua Xiao

The paper systematically compares three fusion paradigms for multi-domain RLVR experts: Merge of task vectors, Mix RL on pooled datasets, and multi-teacher on-policy distillation. Average performance differs by at most 1.4 points yet single-benchmark gaps reach 8.6 points, with domain relations visible in task-vector geometry. Practical guidelines emerge: Merge for cheap fusion of existing experts, Mix RL when training from scratch with tuned proportions, and MOPD when preserving domain-specific gains is paramount.