Zijian Zhou

Speculative decoding (SD) accelerates large language model inference by using a smaller draft model to propose draft tokens that are subsequently verified by a larger target model. We introduce MineDraft, a framework that overlaps drafting with verification across two batches of requests, achieving improvements of up to 75% in throughput and 39% in end-to-end latency compared to standard approaches.

We present CORAL, a framework enabling autonomous multi-agent evolution for open-ended discovery tasks. Rather than relying on predetermined exploration strategies, CORAL employs long-running agents that explore, reflect, and collaborate through shared persistent memory, asynchronous multi-agent execution, and heartbeat-based interventions. Across mathematical, algorithmic, and optimization problems, CORAL achieves 3-10 times higher improvement rates with far fewer evaluations than fixed evolutionary search baselines. Notably, four co-evolving agents improve kernel engineering performance from 1363 to 1103 cycles.

Modern language agents must operate over long-horizon, multi-turn interactions, where they retrieve external information, adapt to observations, and answer interdependent queries. Yet, most LLM systems rely on full-context prompting, appending all past turns regardless of their relevance. This leads to unbounded memory growth, increased computational costs, and degraded reasoning performance on out-of-distribution input lengths. We introduce MEM1, an end-to-end reinforcement learning framework that enables agents to operate with constant memory across long multi-turn tasks. At each turn, MEM1 updates a compact shared internal state that jointly supports memory consolidation and reasoning. Experiments across three domains show that MEM1-7B improves performance by 3.5x while reducing memory usage by 3.7x compared to Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task, and generalizes beyond the training horizon.

We propose TETRIS, a novel method that optimizes the total throughput of batch speculative decoding in multi-request settings. Unlike existing methods that optimize for a single request or a group of requests as a whole, TETRIS actively selects the most promising draft tokens (for every request in a batch) to be accepted when verified in parallel, resulting in fewer rejected tokens and hence less wasted computing resources. We show theoretically and empirically that TETRIS outperforms baseline speculative decoding and existing methods that dynamically select draft tokens, leading to a more efficient batch inference in LLMs.

In-context learning (ICL) allows transformer-based language models to quickly learn a specific task with a few "task demonstrations" without updating their parameters. Taking the viewpoint of recent works showing that transformers learn in context by formulating an internal optimizer, we propose an influence function-based attribution technique, DETAIL, that addresses the specific characteristics of ICL. We show how DETAIL can help improve model performance through demonstration reordering and curation, and prove its wide applicability by showing attribution scores obtained on white-box models are transferable to black-box models.

The Shapley value (SV) is adopted in various scenarios in machine learning (ML), including data valuation, agent valuation, and feature attribution, as it satisfies their fairness requirements. However, as exact SVs are infeasible to compute in practice, SV estimates are approximated instead. This approximation step raises an important question: do the SV estimates preserve the fairness guarantees of exact SVs? We generalise Shapley fairness to probably approximate Shapley fairness and propose fidelity score, a metric to measure the variation of SV estimates. We also propose a novel greedy active estimation (GAE) algorithm that achieves a better fairness guarantee than the de facto Monte-Carlo estimation.