Tag
DoubtLess 提出一种免训练的对比解码方法，用一个更小的推理模型引导大模型输出更简洁的思维链，在七个竞赛数学、研究生科学与编程基准上将 Qwen3-32B 的 CoT 缩短 31% 且准确率不变，优于五个推理时基线。
LOCUS is a task-aware low-rank post-training method that reduces output token length in language models while maintaining preference alignment, achieving up to 39.84% reduction on Pythia-2.8B with minimal parameter updates.
TEFM is a framework that reduces token usage for LLMs processing structured data by compressing observations into compact Behavioral Code tokens, while ensuring faithful predictions through a dual-fidelity optimization approach.
Swift-Qwen3.8-27B is UkisAI's reasoning-efficient derivative of Qwen3.8-27B, reducing thinking tokens by 58.3% while maintaining near-identical performance.
Agentic Video in the Gemini API is a new feature that reduces token consumption by up to 88% while increasing quality for processing long videos, enabling tools like Thumio to generate YouTube thumbnails from links.
This paper proposes a neurosymbolic layer for LLMs that improves logical reasoning accuracy by 8.5% without finetuning and reduces token usage by over 50% on long context tasks.
The article introduces MusCoRe, a color-frequency compression protocol that reduces token usage in AI agent conversations by 71.9%, enabling potential local inference on low-resource devices without GPU dependency.
Graft es una herramienta gratuita y open-source que evita que el agente de Claude Code relea el proyecto en cada tarea, reduciendo drásticamente el consumo de tokens y tiempo, y funciona completamente en local.
BLADE is a lightweight framework that dynamically terminates LLM reasoning by expanding probe checkpoints to sentence, self-doubt, and paragraph boundaries, while adaptively selecting informative hidden layers. Experiments on Qwen3 models show near-baseline accuracy with 24.8% token reduction on Qwen3-8B and 15.8% on Qwen3-4B.
Webfetch is a tool that enables local web search for LLM agents, reducing token usage by 87% and cost by 66%, improving efficiency.
This paper empirically evaluates whether reducing tokens in API-based coding agents reduces actual billed cost, finding that prompt-cache traffic dominates cost and token reduction does not reliably lower costs, and can harm task completion.
A/B testing for agent active tool pruning shows: across 121 Terminal Bench tasks, performance slightly improves (+2.48 percentage points), token consumption reduces by 41.7%, and cost decreases by 31.6%.
A tweet criticizes token reduction fads while highlighting Headroom, an open-source tool by a Netflix engineer that compresses LLM payloads locally to reduce costs by up to 95%.
Conduit is a local MCP gateway that reduces token usage by approximately 90%, helping to cut costs for AI applications.
Headroom, an open-source tool from a Netflix engineer, wraps Cursor or Claude in a local proxy to compress payloads, reducing token usage by up to 95% with zero code changes while preserving logic accuracy.
pi-vcc is an open-source tool that provides pure algorithmic conversation compression for the Pi coding agent, achieving 35-99% token reduction without LLM calls, with lossless history search via vcc_recall.
Proposes Reroute, a training-free plug-in for vision-language models that replaces irreversible visual-token pruning with recoverable routing, allowing tokens to re-enter the pipeline later to improve grounding under aggressive token reduction while maintaining VQA performance.
Introduces Efficient Operator Search (EOS), a unified differentiable framework that generalizes token reduction methods (pruning, merging, pooling, adaptive reweighting) into a shared operator space, automatically searching for optimal operator compositions under budget constraints. The method achieves competitive results across benchmarks and reveals consistent operator patterns.
Vulpine is a compiler that transforms human-readable Python code into a compressed macro representation optimized for LLMs, reducing token count by 13.8% on average while enabling exact structural reconstruction.
AQuaUI is a training-free inference-time token reduction method for GUI agent models that uses adaptive quadtrees to reduce spatial redundancy in screenshots, achieving up to 13.22% speedup and 29.52% fewer visual tokens while retaining 99.06% of performance.