arXiv Computation and Language

Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation

The paper introduces DriftBench, a benchmark that tests whether large language models (LLMs) maintain hard constraints while refining scientific ideas over multiple turns. Using 38 research briefs across 24 domains, the study evaluates seven models from five providers on 2,116 runs, finding that iterative pressure increases proposal complexity and reduces constraint adherence, with a high rate of ‘knows‑but‑violates’ instances. The authors also compare judge stability, re‑elicitation effects, and human versus LLM judgments, and release all benchmark materials.

arXiv AI
Sep 12

SemVerBench: Benchmarking LLM Comprehension of Version-Constraint Resolution Semantics

SemVerBench is a benchmark that evaluates how well large language models (LLMs) understand and apply version-constraint resolution semantics, such as determining whether a version satisfies constraints like ^1.2.3 or >=2.0. The study finds that many models struggle with certain corner cases, with GPT‑5.1 performing poorly while Claude and Opus perform much better. The authors suggest that the failures stem from an activation/application gap rather than a lack of knowledge, and recommend that coding agents delegate version resolution to a dedicated resolver tool.

By Qibai Chen, Zeming Liu
arXiv AI
Oct 16

Sci$\Pi$vot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents

Sciπvot is a 248‑case benchmark that evaluates how well large language model (LLM) agents perform scientific judgment during autonomous research. The benchmark is built from six months of real‑world experiments where senior researchers ran LLM agents on 26 projects across six scientific fields and repeatedly corrected the agents’ decisions. In each case, the agent must choose between its original plan and a human‑corrected alternative using only the project record available before the correction, allowing a direct comparison of LLM and human performance under varying reasoning effort and record access.

By Youran Sun, Xingyu Ren, Chugang Yi, Kejia Zhang, Di Wu, Yingkang Cao, Allen Zhou, Jianping Ye, Suying Liu, Jiaxuan Guo, Jianda Du, Weicheng Ye, Kecheng Liu, Hefan Zhang, Haixin Wang, Jie Xiao, Eric Yang, Haizhao Yang
arXiv AI
Aug 28

MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models

MemToC is a controlled benchmark that tests how large language models resolve conflicts between their internal memory and tool outputs. It contains 6,504 episodes built from 542 factual questions, each paired with a model‑generated closed‑book answer and a tool return whose correctness is known, creating four distinct source‑correctness scenarios. Across five 7‑9B open‑weight models, tool responses overwhelmingly dominate closed‑book answers, and only a minority of instruction‑tuned models correctly retain a verified answer when the tool is wrong, while most follow a correct tool or repeat a wrong tool.

By Arseniy Varlamov, Rishat Zinnatullin, Elisei Rykov, Alexander Panchenko, Ilseyar Alimova
arXiv AI
Sep 4

It's the Problem, Not the Path: Budget and Difficulty Confounds in LLM Reasoning Trajectories

The paper investigates whether large language models’ reasoning traces truly contain early, informative signals or merely reflect budget and difficulty confounds. Using a restart‑controlled truncation probe, the authors compare continuation success rates against from‑scratch restart curves across 178 problem‑model pairs, finding that only one case shows prefix‑limited success and that continuing a model’s own prefix generally outperforms restarting. A difficulty‑controlled test and two generation‑free analyses reveal that early internal signals do not carry outcome information beyond a problem‑difficulty baseline, underscoring the need for proper counterfactual controls.

By Yigit Utku Bulut
arXiv Machine Learning
Sep 22

Do Language Models Know Their Own Constraints?

The study investigates whether language models can explicitly report constraints they have learned through post‑training fine‑tuning. Using constrained recipe generation with five banned ingredients, the authors compare supervised fine‑tuning (SFT) and Group Relative Policy Optimization (GRPO) against an untrained baseline on a Constraint Awareness Benchmark. Both fine‑tuning methods increase behavioral compliance from 4% to about 90% but reduce explicit constraint reporting and erode retained third‑person knowledge, with GRPO showing more destructive effects. The results suggest that reward‑based signals may suppress constraints context‑independently, and that models fail to enumerate constraints on request even when they can avoid them internally.

By Arin Agarwal
arXiv Computation and Language
Sep 22

CCTU: A Benchmark for Tool Use under Complex Constraints

CCTU is a new benchmark designed to evaluate large language models (LLMs) on their ability to use tools under complex constraints. It includes 200 test cases that average seven constraint types and 4,700‑token prompts, covering resource, behavior, toolset, and response dimensions. An executable validation module performs step‑level checks, and nine state‑of‑the‑art LLMs were tested, revealing that none exceed a 20% task completion rate when strict constraints are enforced, with frequent violations and limited self‑refinement.

By Junjie Ye, Guoqiang Zhang, Wenjie Fu, Zelin Li, Tao Gui, Qi Zhang, Xuanjing Huang