Recoleta
GitHub
Source note

TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering

Authors
Yuan Sui, Yulin Chen, Yibo Li, Xue Jiang, Yufei He, Yihong Dong, et al.
Source
arXiv
Published
Collection
Software Intelligence
Coding AgentsActivation SteeringAgent DriftCode IntelligenceSoftware Engineering Benchmarks
Open arXivMarkdown source

Summary

TACT steers coding agents away from two step-level failure modes: repeated reasoning over known facts and repeated tool use without using recent evidence. It claims higher solve rates on software-engineering benchmarks without fine-tuning or prompt changes.

Problem

  • Long coding-agent runs can drift over dozens or hundreds of steps, causing wasted reasoning, repeated searches, redundant file reads, and missed fixes.
  • Final task success does not show whether a failed trajectory broke down through overthinking or overacting, so step-level diagnosis matters for reliable coding agents.
  • These failures matter because they reduce resolve rate and increase steps, context use, and tool calls.

Approach

  • The paper labels each agent step as overthinking, overacting, or calibrated using an LLM judge with a rolling verified state of known facts, hypotheses, edits, test outcomes, and repeated actions.
  • It extracts hidden states at the </think> token, the point where reasoning ends and action begins.
  • It builds two mean-difference activation directions: overthinking vs. calibrated and overacting vs. calibrated.
  • It orthogonalizes the overacting direction against the overthinking direction, then applies test-time residual-stream steering along both axes.
  • The steering variants either cap out-of-band projections, apply a fixed additive shift, or apply a gated partial correction when an activation leaves the calibrated band.

Results

  • Hidden states for overthinking, overacting, and calibrated steps are linearly separable along the drift axes, with AUC about 0.9.
  • Across SWE-bench Verified, Terminal-Bench 2.0, and CLAW-Eval, TACT raises average resolve rate by +5.8 percentage points on Qwen3.5-27B.
  • On Gemma-4-26B-A4B-it, TACT raises average resolve rate by +4.8 percentage points across the same benchmark set.
  • TACT cuts steps-to-resolve by up to 26 steps.
  • The LLM judge reports recalls of 84% for overthinking, 82% for overacting, and 87% for calibrated against human consensus; Cohen's kappa against four references is 0.73 to 0.81.

Link