AI Paper Atlas

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

Consolidating Reasoning with Test-Time Learning

Haotian Wu, Zeming Chen, Hao Zhao, Antoine Bosselut

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

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 changedWhereMeasuredEffect
consolidate trajectories into a LoRA via test-time training instead of in-context aggregationQwen3-4B · AIME25, None tieraccuracy: 32.9 -> 69.5Improved
use TTT-Consolidate instead of best aggregation baselineQwen3 · 10 math/STEM benchmarksaccuracy: up to 20 pointsImproved
increase model scale from 1.7B to 4BTTT-Consolidate · Minority tiergap over extended thinkingImproved
add test-time-learning on incorrect trajectoriesTTT-Consolidate · None tier, 10 benchmarksaverage accuracy gain over baseline: +22.6 (1.7B), +10.6 (4B)Improved
add more incorrect trajectories (2 to 16) (ablation)baseline aggregators · None tieraccuracyNo clear effect
add more incorrect trajectories (2 to 16) (ablation)TTT-Consolidate · None tieraccuracy: 15.2 -> 26.5Improved
spend extra compute on more majority-vote rollouts (ablation)Majority Voting · Minority tieraccuracy: 22.7 -> 22.2 -> 22.6No clear effect
consolidate jointly in one LoRA instead of 8 separate LoRAs with voting (ablation)TTT-Consolidate · None tieraccuracy: 10.3 pointsImproved
swap correct/incorrect answers so majority is wrong (ablation)Prompt and RL Aggregation · Majority-reversedaccuracy: drops up to 52 (Prompt) and 64 (RL) pointsHurt
swap correct/incorrect answers so majority is wrong (ablation)TTT-Consolidate · Majority-reversedaccuracy: drop within [-17, -5]Hurt
replace SFT outer loop with RLVR (GRPO) (ablation)TTT-Consolidate Qwen3-1.7B · None tieraverage accuracy: +4.6; GPQA 22.7 -> 40.0Improved
use SFT aggregation with same data in-context instead of TTT (ablation)SFT Aggregation · None tieraccuracyNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy