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

QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning

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 QuestA, which prepends the first p% of a ground-truth solution to hard math prompts during GRPO-style RL, using a two-stage 50%→25% hint curriculum, to reach new 1.5B-scale SOTA on AIME24/25 and HMMT25.

Problem

RLVR on hard reasoning problems yields sparse rewards and slow learning, while training on easy prompts degrades pass@k; the question is how to adapt RL to solve harder problems more effectively and expand reasoning capacity beyond the base model.

Approach

Input-level question augmentation: prepend a partial ground-truth solution (first p% of the solution block) to hard training prompts so RL gets denser rewards, with the hint removed at evaluation and a curriculum that reduces the hint ratio from 50% to 25%.

Key results

• QuestA-Nemotron-1.5B reaches 72.50 on AIME24 (+10.73), 62.29 on AIME25 (+12.79), and 41.67 on HMMT FEB 25 (+10.11) pass@1, averaging 63.26 vs 53.14 for the base Nemotron-1.5B.
• QuestA-DeepScaleR-1.5B raises the average across math benchmarks from 36.28 to 42.78 (+6.50).
• Removing hints at evaluation, QuestA-Nemotron-1.5B w/o hint reaches 60.44 average, and adding hints during training gives 63.26, with the hint version needing about half the steps.

Surprising findings

• Training with partial-solution hints improves performance at evaluation when no hints are provided, including solving problems unsolved at pass@32 by the base model.
• Training on easy prompts reduces pass@k at larger k, while hard prompts improve it but learn slowly.

Limitations

• Gains are shown mainly for 1.5B math models; cross-domain (code, logic) gains are minor (+0.78 average for DeepScaleR).
• Requires ground-truth solutions to construct partial-solution hints, limiting applicability to domains with reference solutions.
• Theoretical analysis uses a simplified tabular softmax setting, not the autoregressive LLM setting used in experiments.

Experimental setup

Datasets: OpenR1-Math-220K OpenMathReasoning AIME24 AIME25 HMMT FEB 25 Olympiad Bench BRUMO25 GPQA Diamond Zebralogic Code Contests Codeforces LCB V5 202410-202502

Benchmarks: AIME24 AIME25 HMMT FEB 25 Olympiad Bench BRUMO25 GPQA Diamond Zebralogic Code Contests Codeforces LCB V5 202410-202502

Metrics: pass@1 (avg@32) pass@k unsolved problem indices at pass@32

Baselines: Nemotron-1.5B DeepScaleR-1.5B DeepSeek-R1-Distill-1.5B Qwen3-1.7B DeepSeek-R1-Distill-32B Qwen3-8B

Builds on: GRPO DAPO AReaL DeepScaleR

Closest papers by meaning

• General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.91)
• Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.91)
• SATURN: SAT-based Reinforcement Learning to Unleash LLMs Reasoning (NeurIPS 2025 · similarity 0.91)
• RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.90)
• T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling (ICML 2025 · similarity 0.90)
• MATH-Beyond: A Benchmark for RL to Expand Beyond the Base Model (ICLR 2026 · similarity 0.90)