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

VeriThinker: Learning to Verify Makes Reasoning Model Efficient

NeurIPS 2025 · 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.

Compresses long chain-of-thought in large reasoning models by fine-tuning them with LoRA on an auxiliary binary CoT-verification task (Supervised Verification Fine-Tuning) instead of synthetic short-CoT targets, and reuses the verifier for solution-wise speculative reasoning.

Problem

Large reasoning models overthink, producing redundant self-verification steps that inflate inference cost; existing compression methods rely on costly synthetic concise CoT targets via SFT or RL.

Approach

Instead of training on concise target reasoning chains, fine-tune the LRM only on a binary verification task (judge whether a question-solution pair is correct), with loss on verdict tokens only; this sharpens the model's decision on when to self-reflect, reducing tokens while preserving accuracy.

Key results

• On DeepSeek-R1-Distill-Qwen-7B, VeriThinker reduces MATH500 tokens from 3791 to 2125 while raising accuracy from 94.0% to 94.8%.
• On AIME 2025 with R1-Distill-Qwen-7B, tokens drop from 14321 to 10287 and accuracy rises from 38.7% to 40.8%.
• Solution-wise speculative reasoning with Qwen-2.5-Math-7B-Instruct as draft gives 4.2x throughput on MATH500 for R1-Distill-Qwen-7B, activating the LRM on only 14.6% of problems.

Measured effects reported in the paper

Surprising findings

• Verification on reversed or random labels behaves differently: a random-label variant failed to compress CoT, while semantically neutral 'North/South' labels still compressed, suggesting the mechanism is contrastive discrimination rather than learned correctness semantics.
• VeriThinker does not effectively compress the 1.5B reasoning model, as the auxiliary task causes catastrophic forgetting in small models.

Limitations

• Does not effectively compress smaller reasoning models such as a 1.5B model, likely due to catastrophic forgetting.
• Compression is less pronounced on the R1-Distill-Llama-8B model (about 24% token reduction).
• Speculative reasoning throughput benefits may diminish as problem difficulty increases.

Experimental setup

Datasets: PRM12K GSM8K LIMO Numina-Math MATH500 AIME 2024 AIME 2025

Benchmarks: MATH500 AIME 2024 AIME 2025 GSM8K

Metrics: accuracy reasoning tokens verification accuracy precision recall F1 throughput AcR

Baselines: Truncation Fast-Prompting CoT-Valve SFT

Builds on: chain-of-thought speculative decoding LoRA

Closest papers by meaning

• TrimR: Verifier-based Training-Free Thinking Trimming for Efficient Test-Time Scaling (ICLR 2026 · similarity 0.91)
• SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model Reasoning (ICML 2026 · similarity 0.90)
• Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression (ICML 2026 · similarity 0.90)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.90)
• Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.90)
• ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning (NeurIPS 2025 · similarity 0.90)