Papers by Martin Tutek
CRISP: Persistent Concept Unlearning via Sparse Autoencoders (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work has explored sparse autoencoders (SAEs) to perform precise interventions on monosemantic features, but most SAE-based methods operate at inference time, which does not create persistent changes in the model’s parameters. |
| Approach: | They propose a parameter-efficient method for persistent concept unlearning using SAEs that automatically identifies salient SAE features across multiple layers and suppresses their activations. |
| Outcome: | The proposed method outperforms previous methods on safety-critical unlearning tasks from the WMDP benchmark, successfully removing harmful knowledge while preserving general and in-domain capabilities. |
Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps (2025.emnlp-main)
Copied to clipboard
| Challenge: | Language models (LMs) produce a chain of thought (CoT) when prompted to think step-by-step, but it is unclear whether the reasoning encoded in the CoT is faithful to the models’ parametric beliefs. |
| Approach: | They propose a framework for measuring parametric faithfulness of generated reasoning by unlearning reasoning steps (FUR) they propose to erase information contained in reasoning steps from model parameters and measure faithfulness as the resulting effect on the model’s prediction. |
| Outcome: | The proposed framework erases information contained in reasoning steps from model parameters and measures faithfulness as the resulting effect on the model’s prediction. |
REVS: Unlearning Sensitive Information in Language Models via Rank Editing in the Vocabulary Space (2025.findings-acl)
Copied to clipboard
| Challenge: | Current approaches to address this issue involve costly dataset scrubbing or model filtering through unlearning and model editing. |
| Approach: | They propose a method for unlearning sensitive information from language models . they curate email and URL datasets and a social security number dataset . |
| Outcome: | The proposed method shows superior performance and robustness to extraction attacks on real-world datasets. |
Easy to Decide, Hard to Agree: Reducing Disagreements Between Saliency Methods (2023.findings-acl)
Copied to clipboard
| Challenge: | A popular approach to unveiling the black box of neural NLP models is to leverage saliency methods, which assign scalar importance scores to each input component. |
| Approach: | They propose to use saliency methods to evaluate whether an explanation is faithful and argue that Pearson-r is a better-suited alternative to rank correlation. |
| Outcome: | The proposed methods exhibit weak rank correlations even when applied to the same model instance and advocated for alternative diagnostic methods. |
CATfOOD: Counterfactual Augmented Training for Improving Out-of-Domain Performance and Calibration (2024.eacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown remarkable generalization capabilities, performing well on various tasks such as question answering (QA), complex reasoning, and code generation. |
| Approach: | They propose to augment training data of smaller language models with automatically generated counterfactuals (CF) instances to improve out-of-domain (OOD) performance of SLMs in extractive question answering setup. |
| Outcome: | The proposed approach improves out-of-domain (OOD) performance of small language models in extractive question answering setup. |
Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent prompting techniques have improved LLMs’ performance on various reasoning tasks, but there is little understanding of what triggers reasoning abilities in LLM in the inference stage. |
| Approach: | They propose a method that transforms a natural language problem into code and directly prompts the LLM using the generated code without resorting to external code execution. |
| Outcome: | The proposed method boosts multiple LLMs by 22.52 percentage points on GPT 3.5, 7.75 on Mixtral, and 16.78 on Mistral. |
Sequence Repetition Enhances Token Embeddings and Improves Sequence Labeling with Decoder-only Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern language models (LMs) are trained in autoregressive manner, conditioned on the prefix. sequence labeling (SL) tasks assign labels to each individual input token, naturally benefiting from bidirectional context. |
| Approach: | They explore sequence repetition (SR) as a less invasive alternative to decoder-only models . they show that increasing the number of repetitions does not degrade SL performance . |
| Outcome: | The proposed technique improves the quality of token-level embeddings and surpasses encoders and unmasked decoders. |