Daily Brief

The Task-Aware Prompt Rewriter (TAPR) enhances LLM performance by reformulating prompts for tasks like question answering and summarization. Trained with reinforcement learning, TAPR shows consistent improvements over base models, achieving higher accuracy on benchmarks such as Natural Questions and GSM8K. The code is available for further exploration.

The authors introduce Step-Aware Reasoning Energy (SARE), a framework that quantifies computational effort in chain-of-thought reasoning for LLMs, revealing non-uniform energy distribution across reasoning steps. Their findings indicate that incorrect trajectories exhibit lower energy at critical junctions, and SARE features outperform traditional output-based confidence metrics across six benchmarks and three open-weight LLMs.

Cloudflare's Workers AI optimizes GPU memory usage for Kimi K-series and GLM models by quantizing KV caches and compressing model weights, achieving up to 41% higher throughput and 40% reduced memory usage without sacrificing accuracy. This allows for more concurrent requests, enhancing service efficiency and lowering costs for customers.

ReLoop-UME introduces a recurrent model for universal multimodal embedding, achieving 44.9x faster retrieval than UME-R1 and 1.5x faster than PLUME. By utilizing Learnable Retrieval Registers, it enhances retrieval performance on benchmarks MMEB-V2 and MRMR, while maintaining a fixed token workspace.

The paper presents SciDisco, a scalable framework for training Scientific Discovery agents in process-verifiable environments, achieving state-of-the-art results on hypothesis-driven scientific data analysis benchmarks with the SciDisco-14B model. This framework enhances long-horizon scientific analysis by compiling hypotheses, datasets, and verifiers into interactive environments, enabling turn-level credit assignment for actions that yield verifiable evidence.

MMShopBench is a new benchmark for multimodal, multi-turn shopping agents, addressing the limitations of existing text-only benchmarks. It features real shopping logs with annotated purchase intents, enabling agents to infer requirements from images and dialogues. Initial evaluations show that fine-tuning open-source models can significantly reduce performance gaps with proprietary models.

This study evaluates LLMs' ability to predict item difficulty levels in assessments, finding that GPT-4.1 achieved a QWK of 0.578, but was outperformed by ConvBERT at 0.625. The results indicate LLMs struggle with labeling hard items, particularly GPT-5.4, which tends to underestimate difficulty, suggesting caution in using LLMs for targeted item generation.

Self-Supervised Skill Optimization (SSO) introduces a framework for optimizing skills in large language models (LLMs) without ground-truth feedback. By leveraging unlabeled task instances, SSO outperforms existing prompt optimizers in both closed-ended and open-ended tasks, achieving results comparable to GT-based optimizers. This method enhances the reusability of procedural guidance for LLM agents.