Publications
COLA, accepted at ICML 2026, is listed first. This page includes published work, accepted papers, public preprints, and manuscripts currently under review.
CAS quartiles follow the 2025 CAS upgraded list (major category); JCR quartiles follow the 2025 JCR labels. Selected conference venues also show CCF ranks, and impact factors follow official venue metrics available as of March 2026.
Global Credit Assignment via Dynamical Criticality
This ICML 2026 paper studies temporal credit assignment in recurrent, convolutional recurrent, and spiking systems through criticality-driven online local learning with constant activation memory.
- Studies temporal credit assignment beyond BPTT through a criticality-driven local learning rule.
- Uses long-range spatial inertia and temporal correlations to approximate global credit propagation with a locally computable teaching signal.
- Derives a closed-form update requiring constant activation memory and O(H) auxiliary state, and validates accuracy, stability, and scalability on RNN, ConvRNN, and SNN benchmarks.
Decision-Making Large Language Model for Wireless Communication: A Comprehensive Survey on Key Techniques
This survey organizes decision-making LLMs for wireless communication across data construction, architecture adaptation, reasoning control, multi-agent coordination, and open challenges.
- Surveys LLM-enabled decision-making for wireless communication across data, modeling, reasoning, inference, learning-based control, and multi-agent settings.
- I independently contributed the sections on data generation and augmentation, architecture adaptation, and open challenges.
- The paper was published in IEEE Communications Surveys & Tutorials.
Cooperative Edge Caching with Large Language Model in Wireless Networks
We formulate cooperative multi-base-station edge caching as an LLM-native sequential decision problem, build an SFT+GRPO training pipeline, and design an opportunity-aware reward.
- Builds a multi-BS, multi-user cooperative edge-caching environment and formulates it as an LLM-native sequential decision problem.
- Implements a two-stage SFT+GRPO training pipeline with TRL, Unsloth, and QLoRA under a strictly validated text-to-action interface.
- Achieves strong generalization across user scale, base-station count, and request distributions.
LinearARD: Linear-Memory Attention Distillation for RoPE Restoration
LinearARD aligns row-wise dense self-relations between native-RoPE teachers and RoPE-scaled students, restoring long-context performance with an exact linear-memory distillation objective.
- Proposes a restoration-distillation framework that aligns row-wise dense self-relations between a native-RoPE teacher and a RoPE-scaled student.
- Derives an exact linear-memory KL kernel, avoiding explicit construction of full relation matrices on long sequences.
- On LLaMA2-7B scaled from 4K to 32K, the method recovers 98.3% of short-context performance using only 4.25M training tokens.
Cluster-Aware Attention-Based Deep Reinforcement Learning for Pickup and Delivery Problems
CAADRL combines global self-attention, intra-cluster attention, and a dynamic dual-decoder to exploit clustered structure in pickup and delivery problems while reducing inference latency.
- Proposes CAADRL, a cluster-aware deep reinforcement learning framework for pickup and delivery problems with clustered structure.
- Integrates global self-attention, intra-cluster attention, and a dynamic dual-decoder with a learnable gate.
- Achieves leading performance on clustered benchmarks while reducing inference latency relative to neural collaborative-search baselines.