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

Measuring and Mitigating Post-Hoc Rationalization in Reverse Chain-of-Thought Generation

ICML 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.

Measures post-hoc rationalization in reverse chain-of-thought generation with a three-level anchoring hierarchy, shows semantic suppression increases trajectory anchoring, and proposes SSR (skeleton-guided reasoning) and its distilled variant SSR-D.

Problem

Reverse chain-of-thought generation (RCG) builds reasoning traces from query-answer pairs with answer visibility, which yields post-hoc rationalizations that create a train-inference mismatch for students who must reason without the answer.

Approach

Three-level anchoring metrics (lexical, trajectory, probabilistic) quantify answer dependence; semantic suppression is shown to raise trajectory anchoring, and SSR replaces suppression with structural decoupling by first generating an abstract functional-tag skeleton that guides full reasoning, distilled into the target model as SSR-D.</natural_language_placeholder>

Key results

• SSR-D reaches 59.5 on ArenaHard for Qwen3-8B-Think vs 50.8 for NEU, and 50.8 vs 41.7 for Qwen3-4B-Thinking-2507.
• SSR-D on AIME 2025 reaches 44.2 for Qwen3-8B-Think vs 33.3 for NEU and 70.0 untrained.
• SSR lowers probabilistic anchoring display score from 52.3 (NEU) to 40.8, while AUG-SUP raises trajectory anchoring from 47.5 to 52.7.

Surprising findings

• Semantic suppression reduces lexical anchoring but increases trajectory anchoring, so lexical overlap is a misleading indicator of answer dependence.
• NEU-trained Qwen3-8B drops from 70.0 to 33.3 on AIME 2025, below its untrained baseline.

Limitations

• Trajectory and probabilistic anchoring rely on a scorer model and fixed-prefix estimators; metrics are display-normalized against constructed references.
• Path diversity measurements are heavy-tailed and treated only as a stress diagnostic.
• Training data are LMArena prompts with synthetic reference answers, not expert-verified data.

Experimental setup

Datasets: LMArena

Benchmarks: ArenaHard-v2.0 EQ-Bench 3 IFEval MultiChallenge GPQA-Diamond AIME 2025

Metrics: accuracy lexical anchoring trajectory anchoring probabilistic anchoring alignment rate

Baselines: NEU SUP AUG-SUP Best-of-3 rollout

Builds on: chain-of-thought Ironic Process Theory

Closest papers by meaning

• Continuous Chain of Thought Enables Parallel Exploration and Reasoning (ICLR 2026 · similarity 0.87)
• Probing the Trajectories of Reasoning Traces in Large Language Models (NeurIPS 2026 · similarity 0.87)
• Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.87)
• Chain-of-Knowledge: Grounding Large Language Models via Dynamic Knowledge Adapting over Heterogeneous Sources (ICLR 2024 · similarity 0.87)
• Metastable Dynamics of Chain-of-Thought Reasoning: Provable Benefits of Search, RL and Distillation (ICML 2025 · similarity 0.86)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.86)