Research

The problems I want to solve, where the work lands, and the papers.

  1. 01

    LLMs are memorizing, not generalizing enough

    They lean on memorized patterns instead of transferable, compositional structure. How do we train for genuine generalization?

  2. 02

    LLMs have no cerebellum: no fast loop to act and verify in real time

    Learning happens in slow, offline training runs, and every action waits on the whole model. We need to model the cerebellum: a fast forward model that predicts the next step, carries the routine ones itself, and skips the brain loop entirely, waking the slow model only when its prediction misses. That is what a real-time agent needs, most of all a computer-use agent, where every click waits on a round trip through the big model.

  3. 03

    LLMs cannot learn from a stream without forgetting

    Experience arrives once, in order, and never comes back. Learning from it needs a stronger state store than a context window, and learning without backprop that consolidates what matters into memory as the stream passes, rather than waiting for the next offline training run.

  4. 04

    Safe AI cannot be one authoritarian model: it needs a democratic multi-agent society

    A single large model that holds all the power is a single point of failure. Safety should come from a society of agents: many subagents, each with its own safeguards, and a balance of power strong enough that when a share of the agents break, the rest can repair the society, as human societies do. The levers of self-change stay out of reach: reward systems and training pathways are controlled and closed to the agents, and copying weights is banned. Each agent’s freedom and resources are allocated by the society’s consensus rules and laws.

  5. 05

    LLMs decode by classification, not true generation

    Every token is chosen by scoring it against a vocabulary of hundreds of thousands, and that classifier head is the bottleneck for both generation latency and the long tail, where rare tokens are starved of probability. A human does not compare hundreds of thousands of words before producing each one: the activation drives the actuator directly. Generation should work the same way: a truly generative process that emits the output, not a discriminative one that ranks every candidate at every step.

papers

8
  1. 2026in submission · ACM IUI 2027safetypost-training

    Do Not Be Angry at Your Agent: Insights from Human-AI Collaborative Problem Solving

    When an agent gets it wrong over a long multi-turn trajectory, should the human restart it, steer it neutrally, or steer it emotionally? Across twelve models and five benchmarks the best strategy depends on the agent (restarting wins for instruction-tuned models, emotional steering pays for reasoning-specialised ones), but one constraint survives: anger is never the right choice. A Jacobian-lens readout shows why, angry feedback pulls the deep-layer workspace toward apology and agreement and away from task content, turning a correction into sycophantic capitulation.

  2. 2026in submission · ARR 2026 Augustsafetysynthetic data

    Should We Type or Talk to LLM Agents? A Comprehensive Study of Voice and Keyboard Input Perturbations

    Typing and speaking leave different signatures on a prompt: orthographic noise from keyboards, disfluency and restructuring from transcription. HIVE perturbs both channels to measure the cost. Voice perturbations lower accuracy on every instruction-tuned model tested and it is the structure of the transcription, not its fillers, that carries the cost; keyboard noise costs far less. Both trace to one cause: how many of the question’s tokens survive. A thinking budget recovers the keyboard channel almost entirely but leaves speech untouched.

  3. 2026in submission · NeurIPS 2026memorypost-training

    SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion

    Most LLM unlearning methods need a curated retain set to stop general utility collapsing. SHRED needs only the forget set: the highest-information tokens in each instance become forget positions, the rest stay as benign anchors, and one top-k KL self-distillation objective demotes memorized logits while preserving the distribution elsewhere. Pareto-optimal forget/utility trade-off on four benchmarks, robust to relearning and membership-inference attacks.

  4. 2026EMNLP 2026safetypost-training

    Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM

    Persona effectiveness is task-type dependent: expert prompts consistently improve alignment-dependent tasks (safety, preference) but reliably damage pretraining-dependent knowledge retrieval. PRISM teaches models when to invoke a persona via intent-based self-modeling, preserving accuracy while keeping alignment gains.

  5. 2026in progress · aiming ICLR 2027latencymemory

    AttendTwice: Long-Context Inference via Dynamic Chunk-Level KV-Cache Selection

    Two attention paths. A cheap low-rank path scores which chunks of the KV cache matter for a query; the exact path then attends only over the chunks it selected, so most of the cache is never loaded. Selecting contiguous chunks rather than scattered tokens is what makes the gather fast enough to pay for itself; per-token selection is exact, but its scattered access pattern costs more than it saves.

  6. 2025ACM ICMI 2025synthetic datasafety

    Multimodal Synthetic Data Finetuning and Model Collapse

    Studies how vision-language models degrade when fine-tuned on AI-generated multimodal data. Characterizes the collapse dynamics specific to the multimodal regime and proposes mitigation strategies that preserve diversity across modalities.

  7. 2024preprintpretraininglatency

    Lateralization MLP: A Simple Brain-inspired Architecture for Diffusion

    A brain-inspired MLP architecture with hemispheric lateralization applied to diffusion models. Shows competitive sample quality at reduced parameter count, suggesting structured asymmetry as an inductive bias for generative modeling.

  8. 2024in submission · TMLRpretraininglatency

    Static Key Attention in Vision

    A more efficient attention variant for vision transformers that pre-computes a static key projection, reducing per-token compute while maintaining downstream task performance.

Full list on Google Scholar

academic service

reviewing

  • NeurIPS 2024–2026
  • ICLR 2024–2025
  • ICML 2024–2025

teaching assistant

  • DSCI 552 Machine Learning for Data Science (2023, 2025)
  • CSCI 576 Multimedia Systems Design (2022, 2024, 2025)
  • CSCI 567 Machine Learning (2022, 2024)
  • CSCI 566 Deep Learning and Its Applications (2024)
  • CSCI 544 Applied Natural Language Processing (2023)
Zizhao Huloading