Kang Chen
Ph.D. Student · Fudan University
I work on the intrinsic interpretability of large language models — understanding and improving their reasoning and agentic capabilities.
NAD · ICML 2026 Spotlight ARM · EMNLP 2026 Main TAAR · ACL 2026 Findings
Papers
Sep 2025ICML 2026 Spotlight Do LLMs Signal When They're Right? Neuron Agreement Decoding — selects the answer by internal-activation consensus (Jaccard), reading no answer strings; predicts correctness within 32 tokens and cuts ~98% of the compute. Read → Jan 2026EMNLP 2026 Main Many Expert Agents, One Generalist — No Training Agent-Role Merging (ARM) — training-free consolidation of benchmark-specialist agents: build candidate backbones, select one by an answer-free activation-overlap score, then transplant a few role-salient neurons while protecting the rest. Best average across two model families — beating even the oracle expert selector. Read → Jan 2026ACL 2026 Findings When Does Thinking Longer Stop Helping? Thinking Traps — after an early wrong commitment a model elaborates a self-consistent but incorrect prefix (89% of failures). TAAR is a test-time controller that predicts where the trap is and how deep, truncates before it, and adaptively restarts — +4.6 average on an 8B model with no fine-tuning. Read → Aug 2026 Disagree to Explore, Agree to Commit RISA — Routing-Informed Steering and Arbitration. MoE router traces as a behavioral coordinate for software agents: novelty against recent history steers exploration, guarded peer support gates patch writing, and decision-token agreement across independent attempts picks the final patch — lifting SWE-bench Verified macro-average from 44.9% to 48.2% over uniform on gpt-oss, matching text consensus without reading a line of diff, and transferring to Qwen3.6 (+3.5, p<0.001). Read → Aug 2026 What Is the Model Thinking That It Never Writes Down? JAR — Jacobian Axis Readout. Distils the vocabulary-scale Jacobian lens into J64, a 64-axis named readout of latent reasoning state built with no outcome, effort or difficulty labels; native expert routing reconstructs it at a median per-axis r of 0.69–0.86, keeping 95–100% of its held-out outcome gain for free — and it drives selection, mid-run stopping, and router edits that produce the named behaviour. Read → Jul 2026 When Do Reasoning Chains Diverge — and Regroup? Routing effective rank — a label-free spectral read of an MoE rollout cohort. It traces a reproducible concentration → differentiation → reconcentration arc in 98.5% of 3,105 cohorts, two thirds of it carried by a single cohort-wide common mode, and higher reasoning effort delays the peak by 2.59 octaves. Read → Jun 2026 Does the Same Token Mean the Same State? MoE routing as an answer-string-free agreement signal — RAD selects the densest route basin, reading no answers. Read → May 2026 What Do Agent Benchmarks Actually Test? TraceGraph — pools many models’ runs into one shared decision landscape of Access, Trap, and Repair, exposing process differences a score hides; its trap map lifts SWE-bench Verified resolved rate by up to +3.8 points. Read → May 2026 Does the Same Answer Mean the Same Reasoning? SliceGraph — an activation-key atlas of many CoT runs. In 85.5% of problem-model cells, same-answer correct chains split into more than one process family, and 76.6% of same-answer run pairs are process isomers — same answer, different reasoning route. A measurement object, not a decoder. Read →
Writing
All notes →Short notes and essays — research ideas, talks, and the occasional aesthetic experiment.