Research map / Large language models / Reasoning and chain of thought

STaR: Sensitive Trajectory Regulation for Unlearning in Large Reasoning Models

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 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 STaR, a parameter-free inference-time unlearning framework for large reasoning models that suppresses sensitive content across chain-of-thought trajectories, and introduces MCS and multi-granularity MIA metrics, evaluated on R-TOFU.

Problem

Machine unlearning in large reasoning models must remove sensitive information embedded throughout multi-step chain-of-thought trajectories, not just final answers, and remain robust across decoding strategies such as ZeroThink and LessThink.

Approach

Combines semantic-aware sensitive content detection, a secure prompt prefix, trajectory-level validation of sensitivity and fluency with re-decoding or refusal fallback, and token-level soft/hard logit suppression, all without updating model weights.

Key results

• On forget10, STaR reaches MU 0.95, AFE 0.84, CFE 0.63 and MCS 0.67, the best in each column, versus the best baseline PO at MCS 0.49.
• On forget01, STaR lowers MIA-A AUC to 0.51 versus 0.64 for the best baseline PO.
• Reported running time is 0.5 h for STaR versus 6.0 to 9.0 h for the parameter-tuning baselines.

Surprising findings

• Under paraphrased prompts, baseline AFE drops to roughly 0.31 to 0.37 while STaR retains 0.82.

Limitations

• Relies on an external sensitive-content classifier and LM-based fluency scoring whose quality is not analyzed.
• Evaluation limited to a single benchmark (R-TOFU) and a single 8B backbone.
• The MCS and MIA-C metrics are newly proposed by the authors and their validation is limited.

Experimental setup

Datasets: R-TOFU

Benchmarks: R-TOFU

Metrics: MU AFE CFE MCS Avg (harmonic mean) MIA-A AUC MIA-C AUC ROUGE-L cosine similarity entailment score

Baselines: GA GD KL DPO ECO SPUL

Builds on: R-TOFU gradient ascent direct preference optimization ECO SPUL

Closest papers by meaning

• Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models (CVPR 2026 · similarity 0.87)
• A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.87)
• Textual Unlearning Gives a False Sense of Unlearning (ICML 2025 · similarity 0.86)
• Unified Parameter-Efficient Unlearning for LLMs (ICLR 2025 · similarity 0.86)
• On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.86)
• Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.86)