Layer-wise evaluation improves heterogeneous LLM agent decision making
GLIDE: Generalized Layer-wise Intrinsic Distributional Evaluation for Heterogeneous LLM Agents
Artificial IntelligenceMultiagent Systems
Summary
Large language model (LLM) agents often need to evaluate their reasoning step by step to decide the best next move, but it's hard to do this cheaply and accurately, especially when different agents with varying styles are involved. The authors introduced GLIDE, a method that looks inside the model’s layers to measure how well a step fits with overall changes across the model. GLIDE then adjusts this internal evidence based on the agent’s past scores and gives a cautious value estimate that helps pick the best reasoning path. Tests show GLIDE improves task success and saves computation without extra checking programs or task-specific training.
What this means in practice
- •For ai development teams: Improve step-by-step decision making in multi-agent language AI systems without needing external evaluators.
- •For automated reasoning software builders: Enhance efficiency and reliability in software that chains logical steps by internally assessing partial computations.
Authors
Wei Zhu, Yiming Wang, Rui Wang, Lixing Yu, Kun Yue, Zhiwen Tang
Abstract
LLM agents require reliable step-level evaluation to compare candidate branches and allocate computation effectively. However, lightweight evaluation remains challenging. External verifiers introduce additional inference cost, while agent-produced confidence or self-evaluation scores can be miscalibrated, especially when candidates are generated by heterogeneous agents. We propose \textbf{G}eneralized \textbf{L}ayer-wise \textbf{I}ntrinsic \textbf{D}istributional \textbf{E}valuation (\textbf{GLIDE}) for LLM agents. \textsc{GLIDE} derives intrinsic step evidence from layer-wise residual coherence, which measures whether local residual updates consistently support the global residual change induced by a candidate step. It calibrates this evidence against the recent score distribution of the generating agent and converts it into a pessimistic reward that jointly accounts for absolute residual evidence and agent-relative standing. The reward provides a cross-agent value signal for MCTS branch selection, while normalized predictive uncertainty guides adaptive branching. Experiments on multi-hop reasoning, sequential decision making, and symbolic logic show that \textsc{GLIDE} improves task performance, step-level ranking quality, and computational efficiency without external verifiers or task-specific supervision.