Foundation Models & Reliable Infrastructure
I am actively seeking PhD programs and research / industry positions for Fall 2027. Feel free to reach out!
Highlights
I'm open to collaborations on related projects — feel free to reach out. Papers are sorted by recency; * indicates equal contribution.
WeDLM is a diffusion language model framework built on standard causal attention via Topological Reordering, enabling prefix-cache compatibility and streaming parallel decoding. It achieves up to 3× speedup on reasoning benchmarks and 10× in low-entropy regimes compared to vLLM-served AR baselines—the first DLLM to outperform industrial AR engines in wall-clock speed.
We propose Bifrost, a self-supervised contrastive learning method that teaches PLMs to capture multi-level fallibility representations in logs, outperforming existing PLMs on anomaly detection, root cause localization, and fault identification.
We introduce Near-Free Parallelism (NFP), the maximum decode positions executable at near-free latency, isolating system-side cost from algorithmic speedups. A principle predicting NFP from hardware balance and kernel granularity reveals that idle-compute intuition can over-predict by up to 23× on Dense and MoE models.
We propose Bifrost, a self-supervised contrastive learning method that teaches PLMs to capture multi-level fallibility representations in logs, outperforming existing PLMs on anomaly detection, root cause localization, and fault identification.
WeDLM is a diffusion language model framework built on standard causal attention via Topological Reordering, enabling prefix-cache compatibility and streaming parallel decoding. It achieves up to 3× speedup on reasoning benchmarks and 10× in low-entropy regimes compared to vLLM-served AR baselines—the first DLLM to outperform industrial AR engines in wall-clock speed.
Inspired by how human SREs operate, we introduce RCLAgent, a multi-agent recursion-of-thought framework that accurately localizes the root cause of failures in microservice systems using only a single request.
We propose d-TreeRPO, a reliable RL framework for diffusion LLMs that leverages tree-structured rollouts and bottom-up advantage computation with verifiable rewards, achieving significant gains on reasoning benchmarks.
We propose DUET, a closed-loop framework for joint exploration of user-item textual profiles in recommendation systems. It distills raw data into concise cues, expands them into rich profiles via self-prompt construction, and optimizes profiles jointly with reinforcement learning using downstream recommendation feedback, while enabling interpretable LLM-compatible representations.
VarParser introduces a variable-centric log parsing strategy that leverages variable information to improve parsing accuracy and efficiency through variable contribution sampling, variable-centric parsing cache, and adaptive variable-aware in-context learning.
RuntimeSlicer introduces a task-agnostic unified runtime state representation that jointly encodes metrics, traces, and logs, enabling lightweight state-aware adaptation for failure management tasks without redesigning modality-specific pipelines.
AIMS is a content-aware resource management approach for AI assistant systems that leverages the content of user requests to inform resource allocation decisions, improving serving efficiency over content-agnostic strategies.
EAGER leverages reasoning trace representation and historical failure patterns in multi-agent systems to support efficient step-wise failure detection, diagnosis, and reflexive mitigation.
AMER-RCL employs a Recursive Reasoning RCL engine with multi-agent framework that performs recursive reasoning on each alert to progressively refine candidate causes, while Agentic Memory incrementally accumulates and reuses reasoning from prior alerts to reduce redundant exploration and lower inference latency.
SpecRCA adopts a hypothesize-then-verify paradigm, leveraging a hypothesis drafting module to rapidly generate candidate root causes, and then employing a parallel root cause verifier to efficiently validate them, achieving superior accuracy and efficiency for scalable and interpretable RCA in complex microservice environments.
We propose Support Vector Regularization (SVR) to control optimization trajectory drift in contrastive language-audio pretraining by using an auxiliary support vector to harness rich information from negative samples while improving training stability.
GeneralLog collaborates LLMs and small models at the knowledge level, routing proprietary logs to LLMs and general logs to small models to reach 90%+ F1 under fully zero-label settings.
We propose CoorLog, a framework using an adaptive coordinator for efficient and generalizable log anomaly detection, especially in evolving software systems.
We introduce 'diagnostic faithfulness' as a key metric and propose FaithLog, a system that enhances model trustworthiness via a causality-guided attention mechanism.
We introduce ZeroLog, a framework for zero-label, generalizable cross-system anomaly detection using logs.
CSLParser presents a collaborative framework where small and large language models work together for efficient log parsing.
We propose FreeLog, a system-agnostic meta-learning approach for cross-system log anomaly detection that requires no labeled data from the target system.
We propose VISTA, a variable-centric strategy that improves the efficiency and reduces the cost of LLM-based log parsing through novel sampling, caching, and ICL techniques.
We propose CLSLog, a collaborative scheme combining LLM generalization and small model efficiency to effectively handle evolutionary logs in anomaly detection.
RoboDojo is a unified benchmark for generalist robot manipulation policies spanning 42 simulation tasks and 18 real-world tasks, evaluating generalization, memory, precision, long-horizon execution, and open-vocabulary instruction following through a single integration layer with a reproducible remote real-world evaluation system and public leaderboard.
We present CodeAD, a novel framework that automatically synthesizes lightweight Python rule functions for LogAD using LLMs, achieving 3.6% F1 improvement while processing datasets 4x faster at a fraction of the cost.
We propose WarriorMath, a defect-aware framework that improves LLM mathematical reasoning through targeted data synthesis and progressive training.
FusionLog routes unlabeled target logs by semantic similarity, applies meta-learned small models to general logs, and distills LLM-guided pseudo labels for proprietary logs to fuse knowledge without target labels.
Outside of research, I recharge through travel, fitness, and landscape photography. I've chased auroras at the edge of the earth and I'm already dreaming of the next adventure.
