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

Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training

NeurIPS 2025 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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.

Identifies MoE "cognitive experts" in DeepSeek-R1 and Qwen3-235B via nPMI between expert activations and thinking tokens, then amplifies two of them at inference time (no training) to improve AIME accuracy and token efficiency.

Problem

Large reasoning models suffer from overthinking and underthinking; it is unclear whether specific MoE experts govern meta-level reasoning behavior and whether steering them can improve reasoning without additional training.

Approach

Uses normalized pointwise mutual information between expert activations and thinking markers (<think>, </think>, "Alternatively") to find cognitive experts from a single forward pass, then multiplies their routing weights by a reinforcement factor at inference time.

Key results

• On AIME24, reinforcing two cognitive experts raises DeepSeek-R1 accuracy from 73.3% to 83.3% while reducing tokens from 9,219 to 8,317.
• On unseen AIME25, the AIME24-identified experts raise DeepSeek-R1 from 63.3% to 73.3% and Qwen3-235B from 66.7% to 73.3%.
• Across methods on AIME24/25, the approach averages 78.7% versus 68.3% for vanilla DeepSeek-R1 and 76.7% for the Tip_t baseline.

Measured effects reported in the paper

Surprising findings

• Reinforcing two randomly chosen experts across multipliers 2 to 512 yields minimal performance variation, while large multipliers on identified experts can collapse output (near-zero accuracy at 512 for top-ranked sets).
• Steering with identified experts increases accuracy while not increasing token usage on AIME24 and AIME25 for DeepSeek-R1, though the paper also reports verbosity increases on ArenaHard.

Limitations

• Expert identification relies on nPMI heuristics and the cognitive expert concept is acknowledged as hypothetical without theoretical justification.
• Validation limited to two open MoE reasoning models (DeepSeek-R1 and Qwen3-235B).
• Results on small benchmarks (AIME 30 problems each) with single-run greedy decoding; pass@k gains under top-p sampling are marginal and AIME25 generalization weaker in pass@1.

Experimental setup

Datasets: AIME2024 GPQA Diamond

Benchmarks: AIME2024 AIME2025 GPQA Diamond ArenaHard

Metrics: accuracy token count underthinking score (Thoughts) pass@1 pass@8

Baselines: Prompt before <think> Prompt after <think> Tip_t vanilla DeepSeek-R1 vanilla Qwen3-235B

Builds on: normalized pointwise mutual information Mixture-of-Experts Tip (decoding constraints for efficient reasoning)

Closest papers by meaning

• Controlling Thinking Speed in Reasoning Models (NeurIPS 2025 · similarity 0.89)
• Efficient Reasoning with Balanced Thinking (ICLR 2026 · similarity 0.89)
• Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.89)
• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.89)
• Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · similarity 0.89)
• Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.88)