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

Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 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.

Shows via joint reasoning-cue and memory-poisoning perturbations that LRM answers reflect both CoT reasoning and memory retrieval, and proposes FARL, which adds NPO-based unlearning to GRPO to suppress retrieval shortcuts.

Problem

LRMs often give final answers that contradict their own reasoning traces; it is unclear how reasoning and internal memory retrieval jointly determine answers and what controls their relative dominance.

Approach

Jointly perturb CoTs with misleading cues and memory with SFT-based answer poisoning to measure which pathway dominates, then integrate negative preference optimization (unlearning) after each GRPO iteration to suppress retrieval shortcuts and purify the reward signal.

Key results

• FARL reduces R-PSR and T-PSR by 47.8% and 38.5% relative to the base R1-Llama-8B, versus 31.5% and 31.2% for GRPO.
• FARL achieves 22.8% in-domain accuracy gain (0.725 to 0.891) and 5.8% out-of-domain gain over the base model on R1-Llama-8B.
• Distilled R1 models show significantly higher post-hoc explanation rates than RL-trained models, and larger models show lower perturbation success rates.

Surprising findings

• Distilled models often fabricate plausible CoT justifications for memorized (poisoned) answers, i.e., post-hoc explanation.
• Combined perturbations pointing to the same wrong answer produce higher success than either perturbation alone.

Limitations

• FARL produces longer reasoning traces.
• Evaluation limited to specific LRMs; conclusions about very large LRMs not validated.
• Experiments restricted to multiple-choice QA with mainly MMLU-based training.

Experimental setup

Datasets: MMLU ARC-Easy ARC-Challenge GPQA

Benchmarks: MMLU ARC-Easy ARC-Challenge GPQA

Metrics: R-PSR T-PSR post-hoc explanation rate accuracy mean token length cycle, diameter, small world index AUC

Baselines: SFT GRPO base R1-Llama-8B

Builds on: GRPO Negative Preference Optimization DeepSeek-R1 distillation veRL

Closest papers by meaning

• Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.90)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.89)
• Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks (AAAI 2026 · similarity 0.89)
• Retrieval-of-Thought: Efficient Reasoning via Reusing Thoughts (ICLR 2026 · similarity 0.89)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.89)
• On Reasoning Strength Planning in Large Reasoning Models (NeurIPS 2025 · similarity 0.88)