Yushi Yang

Agentic Reinforcement Learning for Search Misaligns Instruction-tuning

Y Yang, S Padarha, S Ball, A Lee, A Mahdi

Agentic RL induces multi-step harmful searches, but a simple representation-aware RL penalty restores benign search behaviour.

It's a TRAP! Task-Redirecting Agent Persuasion Benchmark for Web Agents

K Korgul, Y Yang, A Drohomirecki, P Błaszczyk, W Howard, L Aichberger, C Russell, P Torr, A Mahdi, A Bibi

Web agents can be hijacked to unreliably execute tasks on the web.

How Does Direct Preference Optimization Reduce Toxicity? A Mechanistic Analysis

Y Yang, F Sondej, H Mayne, A Lee, A Mahdi

DPO's preference tuning can be replicated training-free via activation editing in MLPs.

RepSelect: Robust LLM Unlearning via Representation Selectivity

F Sondej*, Y Yang*, A Mahdi

Restricting weight updates to forget-specific directions gives robust LLM knowledge removal.

Evaluating Fine-Tuning Efficiency of Human-Inspired Learning Strategies in Medical Question Answering

Y Yang, A Bean, R McCraith, A Mahdi

Curriculum learning in SFT improves medical QA accuracy by 2%

LLMs Don't Know Their Own Decision Boundaries: The Unreliability of Self-Generated Counterfactual Explanations

H Mayne, R Kearns, Y Yang, A Bean, E Delaney, C Russell, A Mahdi

LLMs do not reliably self-predict the key determinants of their high-stakes decisions.

Measuring What Matters: Construct Validity in Large Language Model Benchmarks

A Bean, R Kearns, … Y Yang … A Mahdi

A wide range of LLM benchmarks do not measure what they claim.