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

DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching

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 DTS, a training-free decoding framework that sketches a branching reasoning tree at high-varentropy, low-entropy decision tokens and terminates early on the first (or voted) finished trajectories to favor short, reliable reasoning.

Problem

Parallel thinking methods for large reasoning models generate redundant, independent trajectories and select answers only after full completion, leaving the reasoning space poorly explored and selection late and unprincipled.

Approach

Selectively branch the decoding tree only at decision tokens (low entropy, high varentropy) rather than sampling independent full trajectories, and select the earliest completed trajectory (or a majority vote over a small budget of finished ones), motivated by an observed length-accuracy anti-correlation and a GRPO length-bias argument.

Key results

• DTS-Stable raises average accuracy over standard inference by +18.11 points on DS-Distill-Qwen-1.5B and +17.16 on DS-Distill-Qwen-7B across four datasets (abstract reports ~14% average gain).
• DTS-Greedy reduces average repetition rate on DS-Distill-Qwen-1.5B from 17.45% to 2.67%.
• DTS-Stable reaches 58.45% average accuracy with DS-Distill-Qwen-7B, versus 53.27% for Self-Consistency, and DS-Distill-Qwen-1.5B with DTS-Stable is reported to outperform QwQ-32B-Preview at similar PFLOPs.

Surprising findings

• Shorter completed trajectories are more accurate, an anti-correlation between reasoning length and accuracy that motivates selecting the earliest finished trace.
• Max-logit trajectory selection underperforms shortest-trajectory selection in every setting tested.

Limitations

• Evaluated only on four models up to 7B and on math, science QA, and logic benchmarks; generality to larger models is not shown.
• Relies on hand-set thresholds (tau_v=1.5, tau_h=2.5, K=3, 48 active branch cap) that were chosen without a broad sensitivity study.
• Length-accuracy anti-correlation is an empirical observation with a heuristic theoretical argument from GRPO, not a proven result.

Experimental setup

Datasets: AIME24 AIME25 GPQA-Diamond LiveBench-Reasoning

Benchmarks: AIME24 AIME25 GPQA-Diamond LiveBench-Reasoning

Metrics: accuracy improvement over standard inference repetition rate PFLOPs

Baselines: Standard inference Self-Consistency DeepConf-low DeepConf-high Max-logit beam search selection DS-Distill-Llama-70B QwQ-32B-Preview

Builds on: Self-Consistency DeepConf GRPO

Closest papers by meaning

• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.90)
• HyPER: Bridging Exploration and Exploitation for Scalable LLM Reasoning with Hypothesis Path Expansion and Reduction (ICML 2026 · similarity 0.90)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (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 0.88)
• Efficient Thought Space Exploration Through Strategic Intervention (AAAI 2026 · similarity 0.88)