Study Finds Variable Reliability in LLM Self-Explanations Across Model Families
What happened
A research study from Universidade Federal do Rio Grande do Sul revealed that large language models frequently fail to align their highlighted explanations with their actual outputs when underlying text is altered.
The Context
- Plausibility masks unreliability: While models generate persuasive rationales, evaluating them across sufficiency, comprehensiveness, and counterfactuality shows significant variance across model families.
- Architectures diverge sharply: Testing on IMDB and Steam datasets demonstrated that GPT-4o-mini largely sustained behavioral consistency, whereas Gemma3, Granite8B, and DeepSeek variants exhibited notable volatility under targeted interventions.
- Interpretability gains an empirical benchmark: Accepted for presentation at the 2026 International Joint Conference on Neural Networks (IJCNN) in Maastricht, Netherlands, the framework establishes functional auto-consistency as an operational test for algorithmic trust.
The Number
tested—spanning GPT-4o-mini, Gemma3, Granite8B, and DeepSeek variants—demonstrating that scale and lineage directly alter whether a system's generated rationale matches its functional behavior.
Takeaway
Expect future enterprise AI compliance audits to require programmatic perturbation tests rather than surface-level prompt rationales before models deploy into high-stakes workflows.
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
Did the Model Use That? Counterfactual Training Boosts Self-Attribution in Language Models
Researchers published a preprint on arXiv examining whether frontier language models are self-attributive, meaning whether they indicate when their answers are influenced by information supplied alongside a question. The study found that most of elev...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
Why Are Some Emotions Harder for LLMs? Uncovering the Causal Mechanisms of Emotion Inference via Sparse Autoencoders
A preprint posted to arXiv's cs.CL section examines why large language models (LLMs) recognize some emotions more reliably than others, using sparse autoencoders (SAEs) to identify causal emotion features. The study reports that surprise and fear rel...
Machine Learning preprints from arXiv.
"Core ML theory and methods in daily preprints."
— A47 Editor
Do Language Models Trust Their Own Justifications? A Study on Functional Consistency
A new arXiv preprint in the cs.LG category, titled "Do Language Models Trust Their Own Justifications? A Study on Functional Consistency," examines whether the tokens that large language models highlight as decisive in text classification actually dr...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
Understanding Jev-Style System One Models: Does Structure Reduce Selection Bias?
A new arXiv cs.CL preprint, "Understanding Jev-Style System One Models: Does Structure Reduce Selection Bias?", examines whether decision structures modeled on Jev-style System One behavior reduce selection bias in large language models. The study sh...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
High Scores, Hidden Failures: Evaluating Codebook-Based LLM Content Analysis
A new arXiv preprint in the cs.CL category, titled "High Scores, Hidden Failures: Evaluating Codebook-Based LLM Content Analysis," reports a large-scale empirical failure-mode analysis of large language model (LLM) coding pipelines used to label text...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
Detecting Differences Is Not Understanding Structure: Large Language Models Fail at Graph Isomorphism
A new arXiv cs.CL preprint, "Detecting Differences Is Not Understanding Structure: Large Language Models Fail at Graph Isomorphism," reports that large language models achieve near-perfect accuracy at detecting whether two graphs are isomorphic, but ...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
LLMs are not stochastic parrots: Evidence for meaning-mediated abstraction from conlang-like tasks
A new arXiv preprint in the cs.CL category, "LLMs are not stochastic parrots: Evidence for meaning-mediated abstraction from conlang-like tasks," tests the claim that large language models can only perform statistical pattern matching. Researchers ga...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
Veracity labels flip the emotional tone of language models but leave racial disparities intact
An arXiv study audited four large language models (GPT-5.4 mini, NVIDIA Nemotron 3 Nano, Llama 4 Scout and Claude Haiku 4.5) and found that adding a veracity label to fact-checked headlines strongly changed the emotional tone of their simulated affec...
Computation and Language (NLP) preprints.
"Daily stream of NLP research papers and preprints."
— A47 Editor
Free Lunch for LLMs Confidence: The RL Critic Already Knows
A new arXiv study (cs.CL) titled "Free Lunch for LLMs Confidence: The RL Critic Already Knows" examines what confidence training actually accomplishes for large language models. The authors compared self-estimators and external estimators of comparab...