Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning

A new approach called TableCache has been proposed to enhance low latency in Text-to-SQL tasks by precomputing key-value (KV) caches offline while preserving primary foreign key relationships between tables. This method addresses inefficiencies in existing inference engines like SGLang and vLLM, which generate redundant cache copies when processing queries with varying table orders.

The introduction of MemoBrain, an executive memory model for tool-augmented agents, addresses the challenges of long-horizon reasoning in AI frameworks. This model captures salient intermediate states and their logical relations, enhancing the coherence and goal-directedness of reasoning processes.

A new technical paradigm called ExpSeek has been introduced, enhancing web agents' interaction capabilities by enabling proactive experience seeking rather than passive experience injection. This approach utilizes step-level entropy thresholds to optimize intervention timing and tailor-designed experience content, demonstrating significant performance improvements in Qwen3-8B and Qwen3-32B models across various benchmarks.

ToolRM has been introduced as a new family of lightweight reward models specifically designed for tool-use scenarios, addressing the limitations of existing reward models in aligning large language models (LLMs) with human preferences. This development includes a novel pipeline for generating high-quality preference data and a benchmark for evaluating these models on tool-calling tasks.

KVzap has been introduced as a fast and adaptive method for key-value (KV) cache pruning in transformer-based language models, addressing the critical inference bottleneck caused by growing context lengths. This method achieves 2-4 times KV cache compression with minimal accuracy loss, demonstrating state-of-the-art performance on the KVpress leaderboard.