Two new co-authored papers have been submitted on (1) multilingual CoT faithfulness; (2) LLM failures under misleading context.

Two new papers have been submitted on (1) SAE-based CoT faithfulness metrics; (2) locally sound yet globally insufficient traces in multi-hop reasoning.

🤖 Will join Amazon as an Applied Scientist Intern and will work on agentic user simulation.

Limitations of Automated Simulatability: LLM Simulators Can Bypass Explanations led by Antonin is accepted at BlackboxNLP @ EMNLP 2026!

💥 Extremely excited to shared that 6 papers, including 3 (co)-first authored ones, have been accepted at EMNLP 2026 about: (1) multilingual counterfactual explanation alignment (Findings); (2) the impact of quantization on self-explanation quality and faithfulness (Findings); (3) iterative feedback-driven counterfactual generation (Findings);(4) polarity-aware evidence grounding (Main); (5) a meta analysis of NLG evaluation (Main); (6) a tool-augmented drop-in ConvXAI layer (Findings). See you in Budapest 🇭🇺!

One paper about the limitation of automated simulatability led by Antonin Poché is in submission.

Three new papers have been submitted on (1) LLM-as-a-Judge circuit discovery; (2) CoT faithfulness meta-evaluation under optimization; (3) multilingual self-explanation alignment.