The paper introduces Circuit Condensation, a post‑training method that prunes low‑attribution edges from large causal graphs and trains a low‑rank adapter to preserve behavior. Across four behaviors and eight models, the condensed circuits are on average 8.1× smaller than the strongest frozen baseline, with reductions up to 316×. Experiments show that weight updates drive the size reduction, and detailed ablations reveal dependencies among remaining edges and a more focused set of heads for indirect object identification.
By Sai Adith Senthil Kumar
arXiv:2607. 18921v1 Announce Type: cross Abstract: Circuit extraction identifies a small set of model components whose presence preserves a target behavior under ablation, and the resulting circuit is often read as the mechanism behind that behavior.
By Yang Sheng, Jie Fu
arXiv:2605. 28860v2 Announce Type: replace-cross Abstract: Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities.
By Jeanmely Rojas Nunez, Viraj Sawant, Nathan Allen, Nomgondalai Amgalanbaatar, Yannis Zongo, Vasu Sharma, Maheep Chaudhary
arXiv:2608.30952v1 Announce Type: new
Abstract: Chemistry questions often demand exact computation and database lookups that a language model cannot supply from its parameters, so it must reach for e...
By Armin Dariani, Sifan Wu, Bang Liu, Entao Yang
The paper investigates mechanistic interpretability, focusing on how automated circuit discovery is evaluated. It shows that the commonly used faithfulness objective can favor circuits that reproduce a model’s behavior poorly, creating an objective-level recovery gap. Experiments on four human-reference tasks and InterpBench reveal that many discovery methods misrank candidate circuits, and that restoring excluded signals can correct most of these misrankings without altering the circuits’ behavior.
By Chuqin Geng, Li Zhang, Haolin Ye, Mark Zhang, Luke Zhang, Xujie Si
arXiv:2609.36813v1 Announce Type: new
Abstract: Large language models (LLMs) exhibit strong general capabilities that mechanistic interpretability has attributed to sparse computational circuits. How...
By Chuanpu Liu, Miao Yu, Yikai Cai, Yuanhe Zhang, Zhenhong Zhou, Li Sun, Zuming Jiang, Yufei Guo
Chemistry questions often demand exact computation and database lookups that a language model cannot supply from its parameters, so it must reach for external tools. Tool use here is a three-part prob...
arXiv:2608. 08536v1 Announce Type: new Abstract: Circuit localization is a mechanistic interpretability task whose goal is to identify a sparse subgraph of a transformer's computation graph sufficient to reproduce a particular behavior.
By Chester Tan, Moritz Lampert, Courtney Maynard, Ankit Ramakrishnan, Tina Eliassi-Rad, Ingo Scholtes
arXiv:2606. 12386v1 Announce Type: cross Abstract: Advancing scientific understanding through mechanistic modeling requires posing the right experimental questions to yield maximally informative data.
By No\'emi \'Eltet\H{o}, Nathaniel D. Daw, Kimberly L. Stachenfeld, Kevin J. Miller
The paper introduces On‑Policy Warmup (OPW), a teacher‑guided training stage where a student agent learns from a teacher on its own interaction trajectories before switching to reinforcement learning with verifiable rewards (RLVR). OPW differs from traditional imitation by focusing on states generated by the student’s own decisions, including imperfect actions and recovery situations. The authors provide a theoretical link between on‑policy reverse‑KL distillation and trajectory‑level distribution matching, showing that, under a competent teacher and low distillation loss, OPW can lower bound initial verifier success and reduce reward‑discovery complexity, thereby accelerating RLVR performance.
By Yitong Qiao, Tiantian He, Lei Liu, Yue Shen, Jian Wang, Jinjie Gu, Zhixuan Chu
arXiv:2606. 25394v1 Announce Type: new Abstract: Finding minimal arithmetic circuits for polynomials over finite fields is a combinatorially hard problem central to algebraic complexity theory.
By Rohan Pandey, Michael Ruofan Zeng, Weikun K. Zhang, Kaijie Jin, Naomi Morato, Archit Ganapule, Bhaumik Mehta, Jarod Alper
arXiv:2607. 23263v1 Announce Type: new Abstract: Deciding whether a trajectory actually fulfills its instruction governs how we measure computer-use agents on long-horizon graphical-user-interface tasks and how we train them with reinforcement learning.
By Yang Wan, Zhenhao Zhang, Jierui Wang, Linchao Zhu