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

Mode-conditioning unlocks superior test-time compute scaling

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling

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 mode-conditioning (ModC), which allocates test-time samples evenly across reasoning modes via specialist models or mode-specific prefixes (or gradient-clustered modes), improving Pass@k scaling on Countdown, NuminaMath/MATH500, and OpenThoughts/AIME 2025.

Problem

Parallel test-time sampling is limited by diversity collapse after SFT/RL, where samples concentrate on a few modes and repeat the same errors, so Pass@k scales poorly.

Approach

Train either separate specialist models per reasoning mode or one model with mode-specific prefix tokens, then explicitly split the sampling budget evenly across modes at test time; modes can be discovered automatically by clustering per-example gradients.

Key results

• On OpenThoughts with Qwen2.5-7B, ModC matches the Pass@1024 of standard training using k=256 samples (~4x efficiency) and raises maximum Pass@k.
• On the Countdown adversarial set, ModC with separate models improves Pass@1024 by up to 20% over standard training.
• On NuminaMath/MATH500, ModC gains up to 10% on Qwen2.5-0.5B and 15% on OLMo2-7B over standard mixed-teacher training; gradient clustering achieves 98.7% F1 at recovering teacher identity.

Surprising findings

• Naively mixing teacher data often underperforms the single best teacher, despite more diverse data.
• Balanced DFS/BFS training data alone does not yield gains over standard training on Countdown.

Limitations

• Assumes modes are known or discovered via gradient clustering; only two modes studied in most experiments.
• Prefix variant may face capacity limits and imperfect control; RL was only applied to the separate-models variant.
• Test-time allocation is fixed and even across modes rather than adaptive.

Experimental setup

Datasets: Countdown synthetic search trajectories NuminaMath OpenThoughts-3

Benchmarks: Countdown natural test set Countdown adversarial test set MATH500 AIME 2025

Metrics: Pass@k Pass@1

Baselines: Standard training (mixed teachers) Best single teacher Pass@k RL Standard RL Random partition ModC Balanced-data standard training

Builds on: Pass@k RL Rademacher random projection gradient similarity Countdown (Gandhi et al., 2024) OpenThoughts-3

Closest papers by meaning

• Scaling Test-Time Compute Without Verification or RL is Suboptimal (ICML 2025 · similarity 0.90)
• Optimizing Test-Time Compute via Meta Reinforcement Finetuning (ICML 2025 · similarity 0.89)
• TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification (ICML 2025 · similarity 0.88)
• What If We Allocate Test-Time Compute Adaptively? (ICML 2026 · similarity 0.88)
• Whatever Remains Must Be True: Filtering Drives Reasoning in LLMs, Shaping Diversity (ICLR 2026 · similarity 0.88)