This article summarizes two high-impact ACL 2026 papers: DRP reduces reasoning token usage by 64% while improving accuracy via skill-aware step pruning, and SAGE doubles skill reuse success rates in agents through sequential rollout and skill-integrated rewards.
The ICML 2026 paper introduces CLSR, a framework that lets multiple LLM agents autonomously create compact, reusable symbolic communication protocols (LSFs), cutting generation tokens by 3–6× while preserving Chain‑of‑Thought accuracy across diverse reasoning benchmarks.
The article analyzes how reward‑shaping techniques can shorten the chain‑of‑thought outputs of Qwen 30‑parameter series models by 20‑40% while preserving or slightly improving performance on AIME‑25 and out‑of‑distribution benchmarks, and it details the experimental design, strategic considerations, and practical insights behind this efficient reasoning approach.
The article identifies the hidden “privilege illusion” that degrades on‑policy distillation when privileged information is injected, and introduces Dual On‑policy Distillation (DOPD), a dynamic two‑stream approach that separates true ability gaps from information gaps, achieving superior performance and stability across LLM and VLM benchmarks.
The paper introduces Heima, a framework that compresses chain‑of‑thought reasoning into a small set of abstract “thinking tokens” for multimodal large models, dramatically reducing generated tokens while preserving inference capability, and provides an adaptive interpreter to reconstruct human‑readable reasoning for analysis.
The paper introduces JTok and its dynamic variant JTok‑M, a token‑indexed parameter scaling method that decouples model capacity from compute, achieving up to 35% compute reduction while delivering consistent performance gains across a wide range of downstream tasks and model sizes.
The paper presents d‑OPSD, the first on‑policy self‑distillation framework for diffusion large language models that eliminates reference solutions and extra teacher models, using only one‑tenth of RL steps while achieving equal or superior reasoning performance and markedly higher training efficiency, as demonstrated on multiple math‑reasoning benchmarks.
SABER introduces a reinforcement‑learning framework that lets large language models dynamically switch among four token‑budgeted reasoning modes, dramatically cutting inference length while preserving or improving accuracy across math, code, and logic tasks.
The SHAPE framework (Stage‑aware Hierarchical Advantage via Potential Estimation) adds a milestone‑based “reasoning tax” to large language model inference, providing step‑wise correctness signals and penalizing verbosity, which yields an average 3% accuracy gain and a 30% reduction in token consumption across multiple math‑reasoning benchmarks.
This roundup presents recent AI breakthroughs, including Robo‑DM’s efficient robot dataset management, DeepKD’s decoupled knowledge‑distillation trainer, a novel informed white‑box attack exposing weaknesses in LLM alignment defenses, the RePPL hallucination detector, Self‑GIVE’s associative reasoning framework, and LLM‑driven RL ensemble methods.
The article examines IBM Research's Abstract‑CoT approach, which replaces verbose natural‑language chain‑of‑thought reasoning with a compact abstract token vocabulary, achieving up to an 11‑fold reduction in inference tokens while maintaining comparable accuracy across math, instruction‑following, and multi‑hop QA benchmarks.
The PIPO (Pair‑In, Pair‑Out) architecture compresses two input tokens into one latent vector and expands each hidden state to emit an extra draft token, halving effective input length and doubling per‑step output, achieving up to +7.15 pass@4 improvement and 2.64× TTFT / 2.07× TPOT speedups on benchmarks such as AIME 2025, GPQA‑Diamond, LiveCodeBench v6 and LongBench v2 with Qwen3.5‑4B/9B backbones.
This roundup reviews ten recent AI papers covering SkillOpt’s treat‑SKILL.md as trainable parameters, compiling whole agent pipelines into model weights, decentralized AI scientist teams, adding a "sleep" consolidation phase to LLMs, interface‑only fixes for frozen agents, reuse‑aware context‑cost strategies, evaluating AI’s ability to forecast scientific breakthroughs, agent aging benchmarks, the trade‑offs of complex harnesses, and multilingual food‑embedding models.
This article breaks down every architectural and training decision behind Gemma‑4—KV sharing, p‑RoPE, per‑layer embeddings, and a dual‑path MoE + dense MLP—while contrasting its efficiency and performance with Qwen‑3 and GLM‑5 across benchmarks, quantization strategies, and RL pipelines.
This article reviews three consecutive works from Alibaba DAMO Academy on compressing and distilling large pretrained language models—AdaBERT, L2A, and Meta‑KD—detailing their motivations, neural‑architecture‑search‑based designs, loss formulations, experimental results, and insights from a Q&A session.
KSA inserts learnable summary tokens to compress KV cache by a factor of eight, enabling accurate long‑context retrieval with far lower memory and compute costs, and it consistently outperforms full‑attention and other hybrid methods on large‑scale benchmarks.
Google’s recent study shows that the length of a model’s token chain is negatively correlated with inference accuracy, and introduces the Deep Thinking Ratio (DTR) metric to identify truly reasoning tokens, enabling the Think@n strategy to halve compute cost without sacrificing performance.
Recent research by Tsinghua and collaborators dissects On‑Policy Distillation for large language models, revealing that higher‑scoring teachers often fail to improve students unless their thinking patterns align, detailing token‑level overlap dynamics, failure cases, and two practical remedies to rescue ineffective distillation.
Prefix Sliding keeps the task prefix and a sliding window of recent tokens while evicting older intermediate tokens from the KV cache, enabling up to three‑fold speedups for long‑chain inference without retraining and extending reinforcement‑learning rollouts beyond 100 k tokens.
This article explains how knowledge distillation enables massive AI models to be compressed and deployed efficiently, covering its principles, classification dimensions, implementation steps, innovative practices at DeepSeek, real‑world applications, and future research directions.
