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

Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning

ICLR 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.

Proposes ThinkMerge, a training-free decoding method that samples K parallel reasoning traces and averages their pre-softmax next-token logits during the answer phase, enabling parallel test-time scaling for open-ended tasks where majority voting is inapplicable.

Problem

Majority voting aggregates complete solutions for close-ended QA but is undefined for open-ended reasoning such as code generation and web-based deep research, so parallel test-time scaling is hard to apply there.

Approach

Run K reasoning traces, then at each answer-phase decoding step average the pre-softmax logits across all K contexts (each conditioned on the same shared answer prefix), sample one token, and append it to all contexts; this is a logit-level product-of-experts-style ensemble within a single reasoning model.

Key results

• On LiveCodeBench hard, pass@1 rises from 20.69 to 28.97 for DeepCoder-14B-Preview (+8.28) and from 24.14 to 31.72 for Qwen3-8B (+7.58).
• On XbenchDeepSearch, WebSailor-7B improves from 37.8 to 48.0 at N=8 (+10.2) and WebSailor-32B from 50.4 to 57.6 at N=8.
• On AIME 2025 with K=4, Qwen3-4B improves from majority voting 68.0 to ThinkMerge 72.0; Prob-Merge drops to 32.7 for R1-Distill-Qwen-7B vs 47.3 for MV.

Surprising findings

• Lowering the answer-phase temperature gives no consistent gain, suggesting thinking-phase diversity is sufficient.
• The shortest-K anti-overthinking bias does not help on code generation and can hurt executability, and the 3B WebSailor agent degrades at N=4 and 8.

Limitations

• Gains on closed-ended tasks are often within noise and MV is generally stronger with shortest-K selection.
• Small agents (WebSailor-3B) degrade with larger N, so benefits require sufficiently capable base models.
• Requires K forward passes per answer token, increasing inference cost; the vLLM v1 one-step implementation was not released at submission.

Experimental setup

Datasets: AIME 2024 AIME 2025 HMMT 2024 HMMT 2025 GPQA Diamond LiveCodeBench v5 GAIA BrowseComp-en BrowseComp-zh XbenchDeepSearch

Benchmarks: AIME 2025 GPQA Diamond LiveCodeBench v5 GAIA BrowseComp-en BrowseComp-zh XbenchDeepSearch

Metrics: accuracy pass@1 pass@N

Baselines: Majority Voting Prob-Merge single-chain decoding WebSailor baseline

Builds on: majority voting product of experts self-consistency

Closest papers by meaning

• ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models (ICML 2026 · similarity 0.86)
• Skeleton-of-Thought: Prompting LLMs for Efficient Parallel Generation (ICLR 2024 · similarity 0.86)
• SpecExit: Accelerating Large Reasoning Model via Speculative Exit (ICML 2026 · similarity 0.85)
• Scaling Speculative Decoding with Lookahead Reasoning (NeurIPS 2025 · similarity 0.85)
• Let Me Think! A Long Chain of Thought Can Be Worth Exponentially Many Short Ones (NeurIPS 2025 · similarity 0.85)
• Scaling Up, Speeding Up: A Benchmark of Speculative Decoding for Efficient LLM Test-Time Scaling (ICLR 2026 · similarity 0.85)