Research map / Multi-agent systems and games / LLM agents

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

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TACT, a training-free activation-steering method that labels coding-agent steps as overthinking, overacting, or calibrated, derives two drift axes from residual-stream activations at the </think> token, and steers rollouts back toward the calibrated region to raise resolve rate.

Problem

Coding agents degrade over long trajectories (agent drift), with overthinking and overacting failure modes that are hard to detect from outcomes and not addressed at the model-internal level.

Approach

Identify linear drift axes (OT-vs-CAL and OA-vs-CAL, Gram-Schmidt orthogonalized) in residual-stream hidden states at the think-end token, using LLM-judge step labels, and apply test-time bidirectional capping, additive, or gated steering along them during rollouts.

Key results

• TACT-Gate raises SWE-bench Verified resolve rate from 67.0 to 73.3 on Qwen3.5-27B (+6.3 pp) and average resolve rate by +5.8 pp.
• On Gemma-4-26B-A4B-it, TACT-Gate improves average resolve rate by +4.8 pp (43.9 to 48.7).
• TACT-Cap reduces steps-to-resolve on SWE-bench Verified by 26% (75.2 to 55.7) on Qwen3.5-27B while adding no LLM calls or retries.

Surprising findings

• OT-vs-CAL axis is most separable at layer 3 (AUC 0.91), suggesting the drift signal is concentrated in early layers.
• TACT-Cap at k=1.0 leaves resolve rate unchanged (67.1 vs 67.0) yet shortens trajectories from 75.2 to 59.8 steps.

Limitations

• Requires open-weight models with accessible activations, so it does not apply to closed API models.
• Axes are fit on coding trajectories; transfer to non-coding agent domains is untested.
• Labels depend on LLM judges (GPT-5.4 default) with CAL precision of 0.79 and substantial but not perfect human agreement (kappa 0.78).

Experimental setup

Datasets: SWE-bench Verified Terminal-Bench 2.0 CLAW-Eval

Benchmarks: SWE-bench Verified Terminal-Bench 2.0 CLAW-Eval

Metrics: resolve rate pass@3 steps-to-resolve

Baselines: Reminder ReCAP AgentReflect AgentDebug Base Model

Builds on: representation engineering mean-difference steering Gram-Schmidt orthogonalization mini-swe-agent assistant axis

Closest papers by meaning

• SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks (NeurIPS 2026 · similarity 0.88)
• EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions (NeurIPS 2026 · similarity 0.87)
• DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.87)
• CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents (NeurIPS 2026 · similarity 0.87)
• DRIFT-BENCH: Diagnosing CoopeRative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction (ICML 2026 · similarity 0.87)
• RefactorBench: Evaluating Stateful Reasoning in Language Agents Through Code (ICLR 2025 · similarity 0.87)