Get Started
Home
Topics
Search
Library
Research questionHow can large language models allocate reasoning computation to preserve accuracy under limited training and inference budgets?Extended reasoning can improve performance on difficult tasks but may consume substantial computation through slow or unnecessary steps. More efficient reasoning can reduce cost but may sacrifice accuracy.
AI
Evaluation & Benchmarks
Inference Optimization
LLM Pretraining & Post-training
Machine Learning
Mechanistic Interpretability
Natural Language Processing
Reasoning
Research Paper
Small / On-device Models
Latest papersRecent research connected to this question, newest first.RecurTrace: Adaptive Latent Reasoning with Loop-Time MemoryEvidence is limited to a controlled MathQA comparison on a shared looped backbone and generation-accuracy comparisons across 0.6B, 1.7B, 4B, and 8B models. The reported design uses loop-time state access and a halting head trained with oracle signals about whether more depth reduces loss; generalization beyond these tasks, backbones, and compute budgets is not established.research paper · Sep 10, 2026Harnessing the Reasoning Economy: A Survey of Efficient Reasoning for Large Language ModelsThe paper surveys causes of reasoning inefficiency, reasoning behaviors, efficiency strategies, and open challenges across post-training and test-time inference. It provides a synthesis of the field rather than evidence from a single proposed method or benchmark.research paper · Sep 8, 2026</think> Doesn't Stop Reasoning: Analysis of Spurious CoT TerminationThe evidence concerns four large reasoning models, five benchmarks, and two training-free early-exit methods that inject an end-of-think token. It reports that increasing attention to the injected token reduces spurious termination and answering-phase length, but does not establish behavior beyond these settings.research paper · Sep 3, 2026CUSUM-Shaped Inference-Time Monitoring and Targeted Re-Decoding for Quantized Small Language Model ReasoningThe study examines mathematical reasoning generation with an inference-time, weight-preserving controller that can trigger targeted re-decoding. Evidence comes from MATH-500 and Omni-MATH evaluations: gains are modest on MATH-500, absent on Omni-MATH, and accompanied by additional sampled tokens.research paper · Sep 2, 2026CHASE: Cache-Hole-Adapted Skip Exit for Looped State-Space Language ModelsThe evidence concerns Looped Mamba and Looped Hybrid Mamba-Transformer models, evaluated on modular-arithmetic and induction tasks, language-model pretraining, downstream benchmarks, and threshold-controlled recurrent-step exits. Reported efficiency gains for the cache-hole-adapted exits occur at the studied scales and when prefill is compute-bound; deeper non-looped models retain a validation-perplexity advantage under strict iso-FLOPs comparisons.research paper · Sep 2, 2026When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in ReasoningThe source analyzes confidence at the thinking-termination boundary, internal attention distributions, and attention across prompt segments in large reasoning models. It reports that later continued thinking is associated with higher perplexity and greater attention to the original question; an attention intervention suppresses explicit thinking but reduces accuracy.research paper · Sep 2, 2026
Related questions
How can large language models cut training and inference costs without materially harming accuracy?How can we quantify and reduce divergent, nonsensical reasoning in large language models?How should multilingual models choose the amount of English context used for reasoning in lower-resource languages?How can LLM serving adapt KV-cache capacity as attention demand changes during long-output reasoning?