Yue Xu

Yijun Liu, Yixuan Wang, Yuzhuang Xu et al. · tsinghua

Large language models (LLMs) utilize key-value (KV) cache to store historical information during sequence processing. The size of KV cache grows linearly as the length of the sequence extends, which seriously affects memory usage and decoding efficiency. Current methods for KV cache eviction typically utilize the last window from the pre-filling phase as queries to compute the KV importance scores for eviction. Although this scheme is simple to implement, it tends to overly focus on local information, potentially leading to the neglect or omission of crucial global information. To mitigate this issue, we propose Judge Q, a novel training method which incorporates a soft token list. This method only tunes the model's embedding layer at a low training cost. By concatenating the soft token list at the end of the input sequence, we train these tokens' attention map to the original input sequence to align with that of the actual decoded tokens. In this way, the queries corresponding to the soft tokens can effectively capture global information and better evaluate the importance of the keys and values within the KV cache, thus maintaining decoding quality when KV cache is evicted. Under the same eviction budget, our method exhibits less performance degradation compared to existing eviction approaches. We validate our approach through experiments conducted on models such as Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3, using benchmarks including LongBench, RULER, and Needle-in-a-Haystack. Results indicate an improvement of approximately 1 point on the LongBench and over 3 points on RULER. This proposed methodology can be seamlessly integrated into existing open-source models with minimal training overhead, thereby enhancing performance in KV cache eviction scenarios.

Yixuan Wang, Shiyu Ji, Yijun Liu et al. · tsinghua

Large language models (LLMs) rely on key-value cache (KV cache) to accelerate decoding by reducing redundant computations. However, the KV cache memory usage grows substantially with longer text sequences, posing challenges for efficient deployment. Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries, especially under tight memory budgets. In this paper, we propose Lookahead Q-Cache (LAQ), a novel eviction framework that generates low-cost pseudo lookahead queries to better approximate the true decoding-stage queries. By using these lookahead queries as the observation window for importance estimation, LAQ achieves more consistent and accurate KV cache eviction aligned with real inference scenarios. Experimental results on LongBench and Needle-in-a-Haystack benchmarks show that LAQ outperforms existing methods across various budget levels, achieving a 1 $\sim$ 4 point improvement on LongBench under limited cache budget. Moreover, LAQ is complementary to existing approaches and can be flexibly combined to yield further improvements.

Yixuan Wang, Yijun Liu, Shiyu ji et al. · tsinghua

Large language models (LLMs) suffer from high inference latency due to the auto-regressive decoding process. Speculative decoding accelerates inference by generating multiple draft tokens using a lightweight model and verifying them in parallel. However, existing verification methods rely heavily on distributional consistency while overlooking semantic correctness, thereby limiting the potential speedup of speculative decoding. While some methods employ additional models for relaxed verification of draft tokens, they often fail to generalize effectively to more diverse or open-domain settings. In this work, we propose Reflective Verification, a training-free and semantics-aware approach that achieves a better trade-off between correctness and efficiency. Specifically, we leverage the inherent reflective capacity of LLMs to semantically assess the correctness of draft tokens in parallel during verification. Using prompt-based probing, we obtain both the original and reflective distributions of draft tokens in a single forward pass. The fusion of these distributions enables semantic-level verification of draft tokens that incorporates both consistency and correctness. Experiments across multiple domain benchmarks and model scales demonstrate that our method significantly increases the acceptance length of draft tokens without compromising model performance. Furthermore, we find that the proposed Reflective Verification is orthogonal to existing statistical verification methods, and their combination yields additional 5$\sim$15\% improvements in decoding speed.