Research map / Large language models / Reasoning and chain of thought
Consolidating Reasoning with Test-Time Learning
Haotian Wu, Zeming Chen, Hao Zhao, Antoine Bosselut
Read the paperCheck your idea against 137,217 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 CORAL, a test-time learning method that consolidates complementary reasoning from multiple parallel reasoning traces to improve aggregation for a query.
Problem
Parallel thinking aggregates multiple reasoning traces, but most aggregation methods treat each trajectory independently and ignore complementary sub-processes across the ensemble.
Approach
Learn from the full ensemble of parallel reasoning trajectories at test time, consolidating their complementary reasoning sub-processes rather than aggregating each trajectory independently.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| consolidate trajectories into a LoRA via test-time training instead of in-context aggregation | Qwen3-4B · AIME25, None tier | accuracy: 32.9 -> 69.5 | Improved |
| use TTT-Consolidate instead of best aggregation baseline | Qwen3 · 10 math/STEM benchmarks | accuracy: up to 20 points | Improved |
| increase model scale from 1.7B to 4B | TTT-Consolidate · Minority tier | gap over extended thinking | Improved |
| add test-time-learning on incorrect trajectories | TTT-Consolidate · None tier, 10 benchmarks | average accuracy gain over baseline: +22.6 (1.7B), +10.6 (4B) | Improved |
| add more incorrect trajectories (2 to 16) (ablation) | baseline aggregators · None tier | accuracy | No clear effect |
| add more incorrect trajectories (2 to 16) (ablation) | TTT-Consolidate · None tier | accuracy: 15.2 -> 26.5 | Improved |
| spend extra compute on more majority-vote rollouts (ablation) | Majority Voting · Minority tier | accuracy: 22.7 -> 22.2 -> 22.6 | No clear effect |
| consolidate jointly in one LoRA instead of 8 separate LoRAs with voting (ablation) | TTT-Consolidate · None tier | accuracy: 10.3 points | Improved |
| swap correct/incorrect answers so majority is wrong (ablation) | Prompt and RL Aggregation · Majority-reversed | accuracy: drops up to 52 (Prompt) and 64 (RL) points | Hurt |
| swap correct/incorrect answers so majority is wrong (ablation) | TTT-Consolidate · Majority-reversed | accuracy: drop within [-17, -5] | Hurt |
| replace SFT outer loop with RLVR (GRPO) (ablation) | TTT-Consolidate Qwen3-1.7B · None tier | average accuracy: +4.6; GPQA 22.7 -> 40.0 | Improved |
| use SFT aggregation with same data in-context instead of TTT (ablation) | SFT Aggregation · None tier | accuracy | No clear effect |
Closest papers by meaning
- DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching (ICML 2026 · similarity 0.88)
- SPRINT: Enabling Interleaved Planning and Parallelized Execution in Reasoning Models (NeurIPS 2025 · similarity 0.88)
- Efficient Reasoning with Balanced Thinking (ICLR 2026 · similarity 0.88)
- TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.88)
- Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization (NeurIPS 2026 · similarity 0.88)
- What If We Allocate Test-Time Compute Adaptively? (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
