Steering LLM Thinking with Budget Guidance (2026.findings-acl)

Copied to clipboard

Challenge: Existing budget control methods for large language models are inadequate for long reasoning . budget guidance can be used to control reasoning length without fine-tuning .
Approach: They propose a budget guidance method that models a Gamma distribution over remaining thinking length during next-token generation and uses it to guide generation in a soft, token-level manner.
Outcome: The proposed method achieves up to 26% accuracy gain on the MATH-500 benchmark compared to baseline methods while maintaining competitive accuracy with only 63% of the thinking tokens used by the full-thinking model.

Similar Papers

Reasoning in Token Economies: Budget-Aware Evaluation of LLM Reasoning Strategies (2024.emnlp-main)

Copied to clipboard

Challenge: Existing evaluations that focus on performance metrics miss a key factor: increased effectiveness due to additional compute.
Approach: They propose to incorporate the compute budget into evaluations to provide a more informative comparison that takes into account both performance metrics and computational cost.
Outcome: The proposed framework outperforms reasoning strategies when they use comparable compute resources.
Token-Budget-Aware LLM Reasoning (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods to enhance reasoning capabilities of large language models incur significant overhead in token usage, leading to increased costs.
Approach: They propose a token-budget-aware LLM reasoning framework that adjusts the number of reasoning tokens based on the reasoning complexity of each problem.
Outcome: The proposed method reduces token costs in CoT reasoning with only a slight performance reduction.
MUR: Momentum Uncertainty guided Reasoning for Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for optimizing reasoning quality are limited by overthinking.
Approach: They propose a method that allocates thinking budgets to critical reasoning steps by tracking and aggregating step-wise uncertainty over time.
Outcome: The proposed method reduces computation by over 45% on average while improving accuracy by 0.33–3.46%.
An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint (2025.emnlp-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) are a powerful tool for test-time scaling, but they are often used under time constraints.
Approach: They propose to use LLMs to make models think before answering questions . they also use self-correction and best-of-N decoding to encourage deeper thinking .
Outcome: The proposed models are able to achieve higher inference accuracy with extra inference computation under time constraints.
Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in large language models (LLMs) have focused on test-time scaling to improve reasoning quality but at the cost of efficiency.
Approach: They propose a training-free framework that enhances reasoning accuracy and stability with minimal overhead.
Outcome: The proposed framework yields consistent gains across general, coding, and STEM tasks while remaining highly efficient.
SelfBudgeter: Adaptive Token Allocation for Efficient LLM Reasoning (2026.findings-acl)

Copied to clipboard

Challenge: Recent large reasoning models have shown exceptional performance on various tasks, but they consume excessive tokens even for simple queries, leading to resource waste and prolonged user latency.
Approach: They propose a self-adaptive reasoning strategy that automatically allocates budgets according to problem complexity and introduces GRPO for reinforcement learning to reduce output length.
Outcome: The proposed model achieves an average response length compression of 61% on math reasoning tasks while maintaining accuracy.
Thinking Economically: A Hierarchical Framework for Adaptive-Complexity Reasoning in LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing efficiency methods for Chain-of-Thought (CoT) generate excessively long rationales without commensurate accuracy gains.
Approach: They propose a training framework that operationalizes this principle through coarse-to-fine budgeting.
Outcome: Experiments on GSM8K and MATH500 show that HAB surpasses standard CoT in accuracy and reduces token usage, achieving stronger performance-efficiency trade-off than baselines.
Can Small Language Models Help Large Language Models Reason Better?: LM-Guided Chain-of-Thought (2024.lrec-main)

Copied to clipboard

Challenge: Existing frameworks for guiding a language model in reasoning tasks are limited by their tendency to generate low-quality rationales that are repetitive and vacuous.
Approach: They propose a framework that leverages a lightweight language model for guiding a black-box large LM in reasoning tasks.
Outcome: The proposed framework outperforms baselines in answer prediction accuracy.
Steering LLM Reasoning Through Bias-Only Adaptation (2025.emnlp-main)

Copied to clipboard

Challenge: Compared with LoRA and BitFit, training a single steering vector per layer with reinforcement learning requires orders of magnitude fewer resources and isolates a much smaller, more interpretable parameter set.
Approach: They propose to train a single steering vector per layer with reinforcement learning while freezing all base weights to match the accuracy of fully RL-tuned reasoning models.
Outcome: The proposed approach improves on an 8 billion-parameter model while keeping all base weights fixed.
Reasoning in Flux: Enhancing Large Language Models Reasoning through Uncertainty-aware Adaptive Guidance (2024.acl-long)

Copied to clipboard

Challenge: Extensive experiments across various reasoning tasks demonstrate that UAG not only enhances the reasoning abilities of LLMs but consistently outperforms several strong baselines with minimal computational overhead.
Approach: They propose an approach to guide LLMs onto an accurate and reliable trajectory by identifying and adjusting uncertainty signals within each step of the reasoning chain.
Outcome: The proposed approach outperforms strong baselines and outperformed strong models with minimal computational overhead.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations