SLT selectively compresses reasoning spans via anticipation and gating, trained in three stages including RL, yielding 22.7% higher accuracy than uniform latent baselines at similar compression and 58.4% shorter chains with 2.8% accuracy drop vs explicit CoT on math benchmarks.
hub Canonical reference
arXiv preprint arXiv:2502.18600 , year=
Canonical reference. 100% of citing Pith papers cite this work as background.
29
Pith papers citing it
Background
100% of classified citations
abstract
Large Language Models (LLMs) have demonstrated remarkable performance in solving complex reasoning tasks through mechanisms like Chain-of-Thought (CoT) prompting, which emphasizes verbose, step-by-step reasoning. However, humans typically employ a more efficient strategy: drafting concise intermediate thoughts that capture only essential information. In this work, we propose Chain of Draft (CoD), a novel paradigm inspired by human cognitive processes, where LLMs generate minimalistic yet informative intermediate reasoning outputs while solving tasks. By reducing verbosity and focusing on critical insights, CoD matches or surpasses CoT in accuracy while using as little as only 7.6% of the tokens, significantly reducing cost and latency across various reasoning tasks. Our code and data are available at https://github.com/sileix/chain-of-draft.
hub tools
citation-role summary
background 5
citation-polarity summary
roles
background 5polarities
background 5representative citing papers
PANDO introduces an online skill-distillation method with a structured library, reflection, demotion, routing, compression, and cache-aware prompting that reaches 58.3% success on 910 VisualWebArena tasks using 58-61% fewer tokens than prior methods.
EGRSD and CL-EGRSD advance the accuracy-length frontier in LLM reasoning by entropy-guided weighting of token-level distillation signals from the teacher.
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
TrigReason matches large reasoning model accuracy on math and science benchmarks by delegating most steps to small models and intervening selectively on three triggers, cutting latency by 43.9% and cost by 73.3%.
Swapping only the orchestration layer around six AI models cut token cost per task 41% and latency 44% with quality at parity, making orchestration a bigger cost lever than model choice.
CLSR lets LLM agents evolve and route symbolic languages that reduce generated tokens by 3-6x versus chain-of-thought while keeping accuracy on benchmarks.
Proposes SCSuff metric for evaluating LLM explanation sufficiency via model-generated alternative inputs, showing explanations are typically insufficient and predictable from hidden states.
DyCon dynamically controls reasoning depth in LRMs by modeling evolving difficulty from step-level embeddings, reducing redundant steps across multiple benchmarks.
Post-hoc model-based compression of reasoning traces cuts training tokens to 12-30% and speeds training 2-7.6x while retaining up to 96% of raw-trace accuracy, though raw traces remain superior at every scale.
High-Entropy Sum (HES) selects high-quality reasoning data for LLMs by summing entropy of the top highest-entropy tokens, matching full-dataset performance with top 20% in SFT and outperforming baselines in RFT and RL.
Proposes nearly balanced TCARDs that minimize the first two generalized word-length pattern components, defines Φ_BCD criterion linked to classical optimality, and constructs designs via coordinate exchange with simulation-calibrated weights for LLM prompt engineering.
PUMA detects reasoning-level semantic redundancy to enable early exit in chains of thought, achieving 26.2% average token reduction across five LRMs and five benchmarks while preserving accuracy and CoT quality.
BET reduces reasoning tokens by about 55% on average while improving performance across benchmarks by learning to short-solve easy queries, fold early on unsolvable ones, and preserve budget for hard solvable queries.
ICR creates a virtual shorter distribution from shortest correct on-policy responses to regularize RL post-training toward concise yet accurate reasoning, improving the accuracy-length Pareto frontier on math and knowledge benchmarks.
On-policy reverse-KL self-distillation from a “be concise” teacher cuts reasoning length up to 56% while preserving or slightly improving math accuracy across several models.
CAT uses intrinsic confidence signals in preference optimization to adapt reasoning length in LRMs, outperforming uniform compression baselines on accuracy across benchmarks.
LASER reduces edge LLM serving latency by 17-38% and improves SLO satisfaction by 3-6% via load-aware adaptive early-exit thresholds and difficulty-aware budget pre-allocation, with 1% average accuracy cost.
SlimSearcher reduces tool-call rounds by 17-58% on GAIA, BrowseComp and XBenchDeepSearch while maintaining accuracy via Pareto filtration in SFT and Adaptive Reward Gating in RL.
HAB applies coarse-to-fine budgeting to LLM reasoning, predicting per-problem depth and learning intra-step token budgets via PPL comparisons and adaptive Pareto optimization, yielding higher accuracy and lower token use than standard CoT on GSM8K and MATH500.
Suppressing anthropomorphic reflection markers via prompt and token interventions preserves or improves LLM reasoning performance on four benchmarks while models continue marker-free verification.
Mixed-Policy Distillation transfers concise reasoning behavior from larger to smaller LLMs by having the teacher compress student-generated trajectories, cutting token usage up to 27% while raising benchmark scores.
SAT reduces reasoning tokens by up to 40% across multiple large reasoning models and benchmarks by adaptively pruning steps based on difficulty while maintaining or improving accuracy.
ES-CoT shortens LLM chain-of-thought generation by tracking runs of identical step answers after linguistic markers, cutting tokens 16% on average while keeping accuracy comparable to full CoT across six datasets and three models.
citing papers explorer
-
Selective Latent Thinking: Adaptive Compression of LLM Reasoning Chains
SLT selectively compresses reasoning spans via anticipation and gating, trained in three stages including RL, yielding 22.7% higher accuracy than uniform latent baselines at similar compression and 58.4% shorter chains with 2.8% accuracy drop vs explicit CoT on math benchmarks.
-
PANDO: Efficient Multimodal AI Agents via Online Skill Distillation
PANDO introduces an online skill-distillation method with a structured library, reflection, demotion, routing, compression, and cache-aware prompting that reaches 58.3% success on 910 VisualWebArena tasks using 58-61% fewer tokens than prior methods.
-
Respecting Self-Uncertainty in On-Policy Self-Distillation for Efficient LLM Reasoning
EGRSD and CL-EGRSD advance the accuracy-length frontier in LLM reasoning by entropy-guided weighting of token-level distillation signals from the teacher.
-
Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
-
TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models
TrigReason matches large reasoning model accuracy on math and science benchmarks by delegating most steps to small models and intervening selectively on three triggers, cutting latency by 43.9% and cost by 73.3%.
-
The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI
Swapping only the orchestration layer around six AI models cut token cost per task 41% and latency 44% with quality at parity, making orchestration a bigger cost lever than model choice.
-
When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning
CLSR lets LLM agents evolve and route symbolic languages that reduce generated tokens by 3-6x versus chain-of-thought while keeping accuracy on benchmarks.
-
What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs
Proposes SCSuff metric for evaluating LLM explanation sufficiency via model-generated alternative inputs, showing explanations are typically insufficient and predictable from hidden states.
-
DyCon: Dynamic Reasoning Control via Evolving Difficulty Modeling
DyCon dynamically controls reasoning depth in LRMs by modeling evolving difficulty from step-level embeddings, reducing redundant steps across multiple benchmarks.
-
Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation
Post-hoc model-based compression of reasoning traces cuts training tokens to 12-30% and speeds training 2-7.6x while retaining up to 96% of raw-trace accuracy, though raw traces remain superior at every scale.
-
Unified Data Selection for LLM Reasoning
High-Entropy Sum (HES) selects high-quality reasoning data for LLMs by summing entropy of the top highest-entropy tokens, matching full-dataset performance with top 20% in SFT and outperforming baselines in RFT and RL.
-
TCARD: Nearly Balanced Two-Level Designs with Treatment Cardinality Constraints with an Application to LLM Prompt Engineering
Proposes nearly balanced TCARDs that minimize the first two generalized word-length pattern components, defines Φ_BCD criterion linked to classical optimality, and constructs designs via coordinate exchange with simulation-calibrated weights for LLM prompt engineering.
-
Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models
PUMA detects reasoning-level semantic redundancy to enable early exit in chains of thought, achieving 26.2% average token reduction across five LRMs and five benchmarks while preserving accuracy and CoT quality.
-
Nice Fold or Hero Call: Learning Budget-Efficient Thinking for Adaptive Reasoning
BET reduces reasoning tokens by about 55% on average while improving performance across benchmarks by learning to short-solve easy queries, fold early on unsolvable ones, and preserve budget for hard solvable queries.
-
Implicit Compression Regularization: Concise Reasoning via Internal Shorter Distributions in RL Post-Training
ICR creates a virtual shorter distribution from shortest correct on-policy responses to regularize RL post-training toward concise yet accurate reasoning, improving the accuracy-length Pareto frontier on math and knowledge benchmarks.
-
CRISP: Compressed Reasoning via Iterative Self-Policy Distillation
On-policy reverse-KL self-distillation from a “be concise” teacher cuts reasoning length up to 56% while preserving or slightly improving math accuracy across several models.
-
CAT: Confidence-Adaptive Thinking for Efficient Reasoning of Large Reasoning Models
CAT uses intrinsic confidence signals in preference optimization to adapt reasoning length in LRMs, outperforming uniform compression baselines on accuracy across benchmarks.
-
LASER: Load-Aware Serving with Early-Exit for Reasoning LLMs at the Edge
LASER reduces edge LLM serving latency by 17-38% and improves SLO satisfaction by 3-6% via load-aware adaptive early-exit thresholds and difficulty-aware budget pre-allocation, with 1% average accuracy cost.
-
SlimSearcher: Training Efficiency-Aware Web Agents via Adaptive Reward Gating
SlimSearcher reduces tool-call rounds by 17-58% on GAIA, BrowseComp and XBenchDeepSearch while maintaining accuracy via Pareto filtration in SFT and Adaptive Reward Gating in RL.
-
Thinking Economically: A Hierarchical Framework for Adaptive-Complexity Reasoning in LLMs
HAB applies coarse-to-fine budgeting to LLM reasoning, predicting per-problem depth and learning intra-step token budgets via PPL comparisons and adaptive Pareto optimization, yielding higher accuracy and lower token use than standard CoT on GSM8K and MATH500.
-
Revisiting Anthropomorphic Reflection Markers in Large Language Model Reasoning
Suppressing anthropomorphic reflection markers via prompt and token interventions preserves or improves LLM reasoning performance on four benchmarks while models continue marker-free verification.
-
Reasoning Compression with Mixed-Policy Distillation
Mixed-Policy Distillation transfers concise reasoning behavior from larger to smaller LLMs by having the teacher compress student-generated trajectories, cutting token usage up to 27% while raising benchmark scores.
-
SAT: Balancing Reasoning Accuracy and Efficiency with Stepwise Adaptive Thinking
SAT reduces reasoning tokens by up to 40% across multiple large reasoning models and benchmarks by adaptively pruning steps based on difficulty while maintaining or improving accuracy.
-
Early Stopping Chain-of-thoughts in Large Language Models
ES-CoT shortens LLM chain-of-thought generation by tracking runs of identical step answers after linguistic markers, cutting tokens 16% on average while keeping accuracy comparable to full CoT across six datasets and three models.
-
Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization
LCPO reduces average LRM output length by over 50% across benchmarks via targeted preference optimization while preserving reasoning performance.
-
Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models
A survey organizing techniques to achieve efficient reasoning in LLMs by shortening chain-of-thought outputs.
-
Measuring AI Reasoning: A Guide for Researchers
Reasoning in language models should be measured by the faithfulness and validity of their multi-step search processes and intermediate traces, not final-answer accuracy.
-
Token-Operations-Oriented Inference Optimization Techniques for Large Models
A survey of large-model inference optimization, organized as a four-layer 'token-operations' taxonomy: multi-model fusion, model optimization, compute-model fusion, and compute-network-model fusion.
- The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modes