Research map / Large language models / Language, safety and interpretability
Directly Optimizing Natural Language Explanations for Behavioral Faithfulness: Simulatability and Recoverability
Advaith Malladi, Shashank Srivastava
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with GRPO on recoverability reward (OPEX-R) instead of prompting base Llama-3.1-8B | Llama-3.1-8B · Explaining classifier predictions (avg) | Simulatability: 0.63 -> 0.86 | Improved |
| train with GRPO on recoverability reward instead of prompting base Llama-3.1-8B | Llama-3.1-8B · Explaining classifier predictions (avg) | Recoverability: 0.66 -> 0.99 | Improved |
| use OPEX-R (8B) instead of larger GPT-4o explainer | GPT-4o · Explaining classifier predictions (avg) | Simulatability: 0.68 -> 0.86 | Improved |
| use simulatability reward (OPEX-S) instead of recoverability reward (OPEX-R) (ablation) | OPEX-R · Explaining classifier predictions (avg) | Simulatability: 0.86 -> 0.86 | No clear effect |
| replace the 8B base with Llama-3.2-3B-Instruct (OPEX-SMALL) (ablation) | OPEX-R · Real datasets (avg) | Simulatability: 0.86 -> 0.85 | No clear effect |
| evaluate with GPT-4o instead of Llama-3.3-70B as User-proxy (ablation) | OPEX-R · Real datasets (avg) | Simulatability: 0.86 -> 0.84 | No clear effect |
| increase labeled input examples from 10 to 25 (ablation) | OPEX · Real datasets | Recoverability and simulatability: recoverability 0.99, simulatability 0.86 at 25 samples | Improved |
| increase labeled input examples from 25 to 100 (ablation) | OPEX · Real datasets | Recoverability and simulatability | Hurt |
| use OPEX instead of chain-of-thought explanations from DeepSeek-R1-Distill-Llama-70B | DeepSeek-R1-Distill-Llama-70B · Real datasets (avg) | Simulatability: 0.66 -> 0.84 | Improved |
| use OPEX instead of chain-of-thought explanations | DeepSeek-R1-Distill-Llama-70B · Real datasets (avg) | Recoverability: 0.73 -> 0.99 | Improved |
| use OPEX-R instead of human-written explanations | Human-written explanations · Car Evaluation | Simulatability: 0.45 -> 0.92 | Improved |
| show OPEX-R explanations to human participants | No explanation · Car Evaluation user study | Human accuracy: 0.63 -> 0.78 | Improved |
Closest papers by meaning
- Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations (ICML 2024 · similarity 0.87)
- ExPO: Unlocking Hard Reasoning with Self-Explanation-Guided Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.87)
- Learning Robust Rationales for Model Explainability: A Guidance-Based Approach (AAAI 2024 · similarity 0.87)
- Probing to Refine: Reinforcement Distillation of LLM Reasoners via Explanatory Inversion (ICLR 2026 · similarity 0.87)
- The Overthinker's DIET: Cutting Token Calories with DIfficulty-AwarE Training (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
