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

What If We Allocate Test-Time Compute Adaptively?

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 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 a training-free, verifier-guided adaptive test-time framework that per problem selects planning, reasoning tools, and compute strategy (BoN/beam/lookahead) over K=10 iterations, using a PRM for intra-iteration pruning and inter-iteration selection, reporting gains on MATH-500, AIME24 and AMO-Bench with Llama-3.1-8B and Qwen-2.5-7B.

Problem

Test-time compute scaling allocates inference compute uniformly with fixed sampling strategies and uses verifiers only for final reranking; the paper asks how to allocate test-time compute adaptively per problem using intermediate verification signals.

Approach

A role-prompted LLM controller (planning, tool selection, compute-strategy and exploration-parameter selection) runs multiple iterations; a math PRM scores step transitions both to prune candidates during generation and to select the final trajectory across iterations, with no learned policy or parameter updates.

Key results

• On MATH-500, Dynamic + PRM Selection raises Llama-3.1-8B from 43.8% (Direct) to 65.4% and Qwen-2.5-7B from 71.2% to 81.4%.
• On AIME24, Dynamic + PRM Selection reaches 10.0% (Llama) and 13.3% (Qwen), versus 3.3% and 6.67% for Direct.
• On AMO-Bench, accuracy doubles from 2.0% to 4.0% for both models under Dynamic (with or without PRM selection).

Surprising findings

• PRM-based selection alone over direct generations gives little or no gain (e.g., AIME24 Llama 3.3% to 6.67%, AMO-Bench unchanged), with the gains coming mainly from adaptive configuration plus PRM selection.
• Fixed configurations reach roughly 81% on MATH-500 with iteration counts, and the adaptive method's advantage over the best fixed configuration is small on MATH-500 (81.4% vs 81.4%).

Limitations

• Effectiveness depends on PRM quality and drops on the hardest problems (MATH-500 Level 5).
• Heuristic prompt-based controller may depend on base model instruction following, especially for smaller models.
• Only mathematical reasoning evaluated; single-run accuracies without variance over seeds.

Experimental setup

Datasets: MATH-500 AIME24 AMO-Bench

Benchmarks: MATH-500 AIME24 AMO-Bench

Metrics: accuracy theoretical FLOPs compute intensity score

Baselines: Direct Direct + PRM Selection Dynamic fixed tool-strategy configurations

Builds on: best-of-N sampling beam search Math-Shepherd

Closest papers by meaning

• Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification (ICML 2025 · similarity 0.90)
• V1: Unifying Generation and Self-Verification for Parallel Reasoners (ICML 2026 · similarity 0.90)
• SCALE: Selective Resource Allocation for Overcoming Performance Bottlenecks in Mathematical Test-time Scaling (AAAI 2026 · similarity 0.90)
• HyPER: Bridging Exploration and Exploitation for Scalable LLM Reasoning with Hypothesis Path Expansion and Reduction (ICML 2026 · similarity 0.89)
• Strategic Scaling of Test-Time Compute: A Bandit Learning Approach (ICLR 2026 · similarity 0.89)
• Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling (NeurIPS 2025 · similarity 0.89)