Tag
The paper introduces DuplexJev, a method that feeds ASR-encoder hidden states through a small connector into a frozen LLM to answer closed-form questions about speech as single-token distributions without decoding, enabling batched decisions about eight utterances in ~0.1s and improving speaker gender/emotion recognition to 90%.
DuplexJev 通过小型连接器将 ASR 编码器隐藏状态输入冻结 LLM，将每个多工决策表示为单一 token 分布而无需自回归解码，使 8 GPU 节点在约 0.1 秒内处理 80 个语音决策，同时在口语问答、说话人性别与情感识别上取得高精度。
The paper introduces Gavel, a method that elicits native skill routing from frozen LLMs via linear projections, enabling efficient tool selection without context overload and outperforming existing pipelines on benchmarks.
This paper introduces Lodestar, a method that uses reinforcement learning to train a short polarizer prompt string that helps a frozen LLM avoid being misled by misleading retrieved passages in RAG question answering. It improves F1 and exact match scores across five QA benchmarks compared to existing entropy-based selection rules.
The paper introduces Hierarchical Self-Improvement (HSI), a framework that enhances frozen LLM agents by evolving task-specific harnesses through hierarchical self-modification, achieving substantial gains on moderate tasks while being limited by feedback quality and backbone capabilities.
IRIS is a training-free framework that uses frozen large language models to construct reusable identity representations for entities in knowledge graphs, enabling efficient entity alignment across different KGs without pair-dependent processing.
GRAB uses a GNN encoder to convert relational tables into latent tokens for frozen LLMs, achieving significant performance gains in multi-table question answering.
This paper introduces PYTHALAB-MERA, an external controller for frozen local LLMs that uses validation-grounded memory and retrieval to improve coding agent performance. It demonstrates superior success rates in strict validation tasks compared to self-refinement baselines by leveraging execution feedback and temporal difference learning.