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

HyPER: Bridging Exploration and Exploitation for Scalable LLM Reasoning with Hypothesis Path Expansion and Reduction

ICML 2026 · 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.

Proposes HyPER, a training-free online expand-reduce controller for multi-path LLM decoding that adaptively chooses among none, single-token MoE routing aggregation, multi-token expansion, and branching, plus a length- and confidence-aware voting rule.

Problem

Test-time scaling with multi-path chain-of-thought needs an exploration-exploitation trade-off under fixed compute; rigid tree search and parallel sampling (self-consistency) handle it poorly, and answer selection by majority or confidence voting is fragile.

Approach

A closed-loop controller scores four actions (None, SingleToken, MultiToken, Branch) from lightweight pool statistics (confidence, entropy, consensus, diversity) every T steps; token-level refinement uses Gumbel-perturbed MoE routing with a two-pass diversity penalty and confidence-weighted logit aggregation; answers are chosen by a length-plus-confidence weighted vote under confidence pruning.

Key results

• On Qwen3-30B, HyPER reaches 94.0% on AIME24 and 95.3% on AIME25 versus 88.0% and 86.0% for SC, at 0.54x and 0.71x token cost.
• On OLMoE, HyPER reaches 86.5% on ARC-E at 0.36x token cost versus 80.1% for SC at 1.00x.
• Length-aware confidence voting improves AIME25 accuracy to 96.7% vs 86.7% for majority voting in the ablation (Qwen3-30B setup).

Surprising findings

• Confidence pruning inverts the typical length bias: surviving correct paths tend to be longer than incorrect ones.
• Confidence-weighted voting can be outvoted by noisy paths, motivating length-aware selection.

Limitations

• Relies on MoE routing diversity for the token-level primitive; dense-model evaluation is only a controller-only preliminary experiment.
• Length signal is only robust when coupled with confidence pruning, not universally valid.
• Hyperparameters (eta, lambda, T, weights) are set by hand and validated by sensitivity analysis only.

Experimental setup

Datasets: AIME24 AIME25 HMMT25 HLE GSM8K MATH500 ARC-C ARC-E

Benchmarks: AIME24 AIME25 HMMT25 HLE GSM8K MATH500 ARC-C ARC-E

Metrics: accuracy normalized token cost

Baselines: SC Self-Certainty DeepConf RoE

Builds on: DeepConf RoE Self-Consistency Tree of Thoughts

Closest papers by meaning

• DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching (ICML 2026 · similarity 0.90)
• ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning (ICML 2026 · similarity 0.89)
• What If We Allocate Test-Time Compute Adaptively? (ICML 2026 · similarity 0.89)
• Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning (NeurIPS 2025 · similarity 0.89)
• Provable Test-Time Scaling for Beam Search in LLM Reasoning (NeurIPS 2026 · similarity 0.89)
• Bag of Tricks for Inference-time Computation of LLM Reasoning (NeurIPS 2025 · similarity 0.89)