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

Learning Robust Rationales for Model Explainability: A Guidance-Based Approach

AAAI 2024 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes G-RAT, a guidance-based selective rationalization method that uses an auxiliary guidance module (noisy weighted scores and prediction distributions) to regularize the selector and predictor, addressing rationalization degeneration and failure.

Problem

Selective rationalization (select-then-predict) suffers from rationalization degeneration (sub-optimal rationales) and rationalization failure (selector encodes class info via positions), producing rationales not aligned with human judgments.

Approach

A guidance module that sees the full input produces a noise-perturbed token weight score and a prediction distribution; the score supervises the selector via cross-entropy (against failure) and the distribution is matched to the rationale predictor via JS divergence (against degeneration), with a linearly decaying coefficient switching the regularizers during training.

Key results

• On the Skew-Selector setting, G-RAT maintains consistent palate F1 up to skew-75 while RNP's palate F1 falls to about 0.5-1.3 at skew-10 to skew-20 in Skew-Predictor.
• On BeerAdvocate, G-RAT achieves F1 of 84.0 (Appearance), 76.6 (Smell) versus FR's 82.8 and 73.4.
• Ablations show removing either L_guide or L_match lowers F1 across all aspects of both datasets.

Measured effects reported in the paper

Surprising findings

• Re-RNP F1 on Skew-Selector palate setting falls to about 0.2 at higher skew thresholds, showing collapse of standard selective rationalization.

Limitations

• Evaluated only with small GRU/GloVe models on review classification; not tested on LLMs.
• Relies on synthetic skew settings to simulate failure, which may not capture all real failure modes.
• Hyperparameters (lambda_guide, lambda_match) chosen empirically with sensitivity analysis on limited ranges.

Experimental setup

Datasets: BeerAdvocate HotelReview

Benchmarks: BeerAdvocate HotelReview Skew-Predictor Skew-Selector

Metrics: token-level precision token-level recall F1 accuracy sparsity S

Baselines: RNP Re-RNP DMR A2R FR

Builds on: RNP A2R FR DMR Gumbel-softmax

Closest papers by meaning

• Tailoring Self-Rationalizers with Multi-Reward Distillation (ICLR 2024 · similarity 0.87)
• Automated Assessment of Fidelity and Interpretability: An Evaluation Framework for Large Language Models’ Explanations (Student Abstract) (AAAI 2024 · similarity 0.87)
• Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery (ICLR 2024 · similarity 0.86)
• A Language Model’s Guide Through Latent Space (ICML 2024 · similarity 0.86)
• Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales? (NeurIPS 2024 · similarity 0.86)
• Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.85)