analiza aiqbench
This research introduces Minimal Test-Time Intervention (MTI), a training-free method that boosts LLM reasoning by selectively applying guidance only to high-uncertainty tokens rather than scaling inference globally. By leveraging the model's existing KV cache for efficient negative-prompt approximation, MTI achieves significant accuracy gains on coding and STEM benchmarks while maintaining low computational overhead. The findings suggest that optimizing test-time behavior through localized uncertainty detection is a more sustainable path forward than indiscriminate compute expansion.
Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention
arXiv:2510.13940v4 Announce Type: replace Abstract: Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized-only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model's KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks-e.g., +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 using Ling-mini-2.0-while remaining highly efficient.
Abstract:Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized-only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model's KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks-e.g., +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 using Ling-mini-2.0-while remaining highly efficient.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2510.13940 [cs.CL]
(or arXiv:2510.13940v4 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 144 cuvinte
Subiecte conexe
Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · 21 iul. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
Import AI 470: No rights for machines; automating environment generation with SPADE; and building better GPU kernels with Hawkeye
Import AI · 24 aug. 2026
Frontier AI Risk Report (February–March 2026)
METR · 19 mai 2026