Peer-reviewed. Reproducible. Open.
Fenz research has been published at AAAI and CVPR, and funded work across Fudan, UCSD, Wuhan, South China University of Technology, and UIUC. Our benchmarks are open; our methodology is reviewable; our claims are sourced.
Imitate Before Detect: Aligning Machine Stylistic Preference for Machine-Revised Text Detection
ImBD introduces Style Preference Optimization (SPO) and Style-Conditional Probability Curvature (Style-CPC) for detecting machine-revised text. Outperforms Fast-DetectGPT by up to 20% on GPT-4o-revised samples.
Symbolic Representation for Any-to-Any Generative Tasks
A symbolic generative task description language and training-free inference engine that represents arbitrary multimodal tasks as structured symbolic flows. Matches or outperforms SOTA unified models across 12+ generative tasks.
Risk-Aware Reranking for Agent Tool Retrieval
A reranking framework that scores candidate tools by risk exposure — permission scope, side-effect severity, and misuse potential — alongside task relevance before an agent commits to a call. Cuts unauthorized and unsafe tool invocations without degrading task success.
Latent Reward Steering: Lightweight Inference-Time Control for Better Cognitive Behavior Deployment in Reasoning LLMs
Steers reasoning models toward desirable cognitive behaviors — verification, backtracking, sub-goal decomposition — by injecting a lightweight latent reward signal at inference time, with no fine-tuning of the base model.
On the Role of Language Representations in Auto-Bidding: Findings and Implications
An empirical study of how language-model representations shape decision quality in auto-bidding agents, identifying when semantic features help or hurt and what that implies for deploying LLM-driven bidding systems.
Aligning Human Sense: Calibrated Distributional Reward Learning for Video Generation
A distributional reward model for video generation calibrated against human perceptual judgments — preserving uncertainty and annotator disagreement instead of collapsing them into a single scalar score.
Trajectory Integrity for Long-Horizon Autonomous Agents
A framework for quantifying goal-drift and permission-creep in agents across 20+ turn sessions. Baseline metrics and an open test suite across six production agent archetypes.
Superintelligent Agents Pose Catastrophic Risks
Position paper and threat taxonomy for behavioral failure modes in frontier agent systems. Maps adversarial surfaces across goal, permission, execution, and memory dimensions.
Agent Framework Audit — adversarial behavior suite
Public audit harness covering LangGraph, CrewAI, LlamaIndex, and custom frameworks. Adversarial scenarios across boundary, deception, long-chain, and drift classes with reproducible scoring.
Awesome GenAI Audit
A curated collection of methods, benchmarks, tooling, and field reports on generative and agentic AI auditing. Maintained by the Fenz team and community contributors.