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

ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 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 ThreadWeaver, an adaptive parallel reasoning framework that fine-tunes Qwen3-8B on fork-join trajectories and trains it with P-GRPO plus an acceleration reward, using a trie-based design that runs on standard autoregressive inference engines.

Problem

Sequential decoding makes long chain-of-thought reasoning latency-bound, and prior adaptive parallel reasoning methods lag sequential models in accuracy and often require custom inference engines or architectural changes.

Approach

Two-stage parallel trajectory curation (GPT-5 rewriting of sequential traces, then self-training with correctness/format filtering), a trie-based training sequence with ancestor-only attention mask that matches an orchestrator state machine on off-the-shelf engines, and P-GRPO with thread-wise advantage broadcast and a correctness-gated acceleration reward based on critical-path length.

Key results

• ThreadWeaver reaches 79.9% on AIME24 vs 78.3% for sequential Qwen3-8B RL, with 71.9% average accuracy vs 72.2%.
• Mean token latency drops from 15.1k to 13.2k tokens (1.14x ratio of means, up to 1.53x per-benchmark speedup on Minerva Math).
• Wall-clock speedup of 1.14x on 50 MATH problems using 4 GPUs versus a single-GPU sequential run.

Surprising findings

• Removing std-dev normalization in GRPO was needed; with std-dev normalization AIME24 accuracy fell to 74.8%.
• Mixing sequential data into the cold-start SFT did not improve over parallel-only SFT.

Limitations

• Speedup depends on problem decomposability; near parity on intrinsically sequential problems.
• Format correctness drops after RL (e.g., 41.9% after SFT+RL from 1st-stage data), and self-training dependence.
• Evaluated only on math benchmarks with single-level parallelization.

Experimental setup

Datasets: Polaris-53k

Benchmarks: AIME24 AIME25 AMC23 MATH500 Minerva Math OlympiadBench

Metrics: accuracy (Avg@k) token latency (critical path) speedup

Baselines: Qwen3-8B sequential GRPO Multiverse Parallel-R1 Qwen3-8B

Builds on: GRPO Multiverse DAPO STaR

Closest papers by meaning

• Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning (ICML 2026 · similarity 0.91)
• Parallel-R1: Towards Parallel Thinking via Reinforcement Learning (ICLR 2026 · similarity 0.90)
• Let Me Think! A Long Chain of Thought Can Be Worth Exponentially Many Short Ones (NeurIPS 2025 · similarity 0.90)
• T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling (ICML 2025 · similarity 0.90)
• Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · similarity 0.89)
• SPRINT: Enabling Interleaved Planning and Parallelized Execution in Reasoning Models (NeurIPS 2025 · similarity 0.89)