AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Directly Optimizing Natural Language Explanations for Behavioral Faithfulness: Simulatability and Recoverability

Advaith Malladi, Shashank Srivastava

ICML 2026 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations

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 changedWhereMeasuredEffect
train with GRPO on recoverability reward (OPEX-R) instead of prompting base Llama-3.1-8BLlama-3.1-8B · Explaining classifier predictions (avg)Simulatability: 0.63 -> 0.86Improved
train with GRPO on recoverability reward instead of prompting base Llama-3.1-8BLlama-3.1-8B · Explaining classifier predictions (avg)Recoverability: 0.66 -> 0.99Improved
use OPEX-R (8B) instead of larger GPT-4o explainerGPT-4o · Explaining classifier predictions (avg)Simulatability: 0.68 -> 0.86Improved
use simulatability reward (OPEX-S) instead of recoverability reward (OPEX-R) (ablation)OPEX-R · Explaining classifier predictions (avg)Simulatability: 0.86 -> 0.86No clear effect
replace the 8B base with Llama-3.2-3B-Instruct (OPEX-SMALL) (ablation)OPEX-R · Real datasets (avg)Simulatability: 0.86 -> 0.85No clear effect
evaluate with GPT-4o instead of Llama-3.3-70B as User-proxy (ablation)OPEX-R · Real datasets (avg)Simulatability: 0.86 -> 0.84No clear effect
increase labeled input examples from 10 to 25 (ablation)OPEX · Real datasetsRecoverability and simulatability: recoverability 0.99, simulatability 0.86 at 25 samplesImproved
increase labeled input examples from 25 to 100 (ablation)OPEX · Real datasetsRecoverability and simulatabilityHurt
use OPEX instead of chain-of-thought explanations from DeepSeek-R1-Distill-Llama-70BDeepSeek-R1-Distill-Llama-70B · Real datasets (avg)Simulatability: 0.66 -> 0.84Improved
use OPEX instead of chain-of-thought explanationsDeepSeek-R1-Distill-Llama-70B · Real datasets (avg)Recoverability: 0.73 -> 0.99Improved
use OPEX-R instead of human-written explanationsHuman-written explanations · Car EvaluationSimulatability: 0.45 -> 0.92Improved
show OPEX-R explanations to human participantsNo explanation · Car Evaluation user studyHuman accuracy: 0.63 -> 0.78Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy