Papers by Adam Mahdi
How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis (2025.emnlp-main)
Copied to clipboard
| Challenge: | Direct Preference Optimization (DPO) is a popular choice of safety fine-tuning algorithms, but prior explanations of its effects only account for dampened toxic neurons in the MLP layers. |
| Approach: | They analysed four language models and found that toxic neurons only account for 2.5% to 24% of DPO’s effects across models. |
| Outcome: | The proposed method outperforms DPO in reducing toxicity while preserving perplexity, without requiring any weight updates. |
LLMs Don’t Know Their Own Decision Boundaries: The Unreliability of Self-Generated Counterfactual Explanations (2025.emnlp-main)
Copied to clipboard
Harry Mayne, Ryan Othniel Kearns, Yushi Yang, Andrew M. Bean, Eoin D. Delaney, Chris Russell, Adam Mahdi
| Challenge: | Existing studies on language models' ability to explain their decisions in natural language have focused on self-generated counterfactual explanations (SCEs). |
| Approach: | They evaluate whether LLMs can generate valid counterfactuals and minimal ones . authors suggest that SCEs are, at best, an ineffective explainability tool . |
| Outcome: | The proposed language models can explain their decisions in natural language, the study finds . the models can produce valid counterfactual explanations, but make small edits that fail to change predictions. |