Large language models (LLMs) have demonstrated exceptional performance across various applications, but their conversational abilities decline sharply as model size decreases, presenting a barrier to their deployment in resource-constrained environments. Knowledge distillation (KD) with Direct Preference Optimization (DPO) has emerged as a promising approach to enhance the conversational abilities of smaller models using a larger teacher model. However, current methods primarily focus on “black-box” KD, which only uses the teacher’s responses, overlooking the rich distributional information within the teacher’s probability distribution. This paper addresses this gap by introducing daDPO (Distillation-Aware DPO), a novel framework that integrates the teacher’s distributional information into DPO distillation while preserving theoretical guarantees. Our framework offers a unified objective that enhances both preference optimization and distribution-based distillation. We provide rigorous theoretical analysis and empirical validation, showing that daDPO outperforms existing methods in restoring performance for pruned models and enhancing smaller models within the same LLM family. Notably, in in-domain evaluation, our method enables a 20% pruned Vicuna1.5-7B to achieve near-teacher performance (-7.3% preference rate), and allows Qwen2.5-1.5B to occasionally outperform its 7b teacher model (14.0% win rate).

KV Cache is commonly used to accelerate LLM inference with long contexts, yet its high memory demand drives the need for cache compression. Existing compression methods, however, are largely heuristic and lack dynamic budget allocation. To address this limitation, we introduce a unified framework for cache compression by minimizing information loss in Transformer residual streams. Building on it, we analyze the layer attention output loss and derive a new metric to compare cache entries across heads, enabling layer-wise compression with dynamic head budgets. Additionally, by contrasting cross-layer information, we also achieve dynamic layer budgets. LAVa is the first unified strategy for cache eviction and dynamic budget allocation that, unlike prior methods, does not rely on training or the combination of multiple strategies. Experiments with four benchmarks (LongBench, Needle-In-A-Haystack, Ruler, and InfiniteBench) demonstrate its superiority over strong baselines. Moreover, our experiments reveal a new insight: dynamic layer budgets are crucial for generation tasks (e.g., code completion), while dynamic head budgets play a key role in extraction tasks (e.g., extractive QA). As a fully dynamic compression method, LAVa consistently maintains top performance across task types.

Large language models (LLMs) possess extensive knowledge and commonsense reasoning capabilities, making them valuable for creating powerful agents. However, existing LLM agent frameworks have not fully utilized past experiences for improvement. This work introduces a new LLM-based agent framework called Retrospex, which addresses this challenge by analyzing past experiences in depth. Unlike previous approaches, Retrospex does not directly integrate experiences into the LLM’s context. Instead, it combines the LLM’s action likelihood with action values estimated by a Reinforcement Learning (RL) Critic, which is trained on past experiences through an offline “retrospection” process. Additionally, Retrospex employs a dynamic action rescoring mechanism that increases the importance of experience-based values for tasks that require more interaction with the environment. We evaluate Retrospex in ScienceWorld, ALFWorld and Webshop environments, demonstrating its advantages over strong baselines.