IC-183Error types in Mistral-7B, Mistral-7B-Instruct, Llama3-8B, and Llama3-8B-Instruct are linearly predictable from internal representations, encoding fine-grained information beyond binary correctness

Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, Yonatan Belinkov

SourceLLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations

The paper defines a taxonomy of five error types based on the distribution of 30 resampled answers: refuses to answer, consistently correct, consistently incorrect, two competing answers, and many answers. Linear probing classifiers trained on intermediate representations of greedy-decoded generations can predict which error type a question will produce. AUC scores range from 0.59 to 0.90 across models and error types on TriviaQA, with the highest scores for 'refuses to answer' (0.85-0.88) and 'many answers' (0.84-0.90). Results on Winobias and Math show similar patterns, confirming that the internal representation encodes more than a single correctness signal.

Evidence

correlational

Key metric

TriviaQA AUC: Mistral-7B (a) 0.86±0.002, (b) 0.88±0.001, (c) 0.59±0.002, (d) 0.63±0.002, (e) 0.90±0.001; Mistral-Instr-7B (a) 0.85±0.011, (b) 0.82±0.008, (c) 0.67±0.002, (d) 0.68±0.006, (e) 0.84±0.003; Llama3-8B (a) 0.87±0.002, (b) 0.86±0.001, (c) 0.59±0.002, (d) 0.61±0.001, (e) 0.89±0.001; Llama3-Instr-8B (a) 0.88±0.014, (b) 0.81±0.002, (c) 0.64±0.003, (d) 0.65±0.004, (e) 0.89±0.001

Caveat

The taxonomy is non-orthogonal; some errors may belong to multiple categories. Performance on error types (c) consistently incorrect and (d) two competing is lower, though still above random. The taxonomy covers 96% of errors in TriviaQA for Mistral-7B-Instruct.

Model

Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral 7B Instruct v0.2, Llama 3 8B, 8B Instruct

Concepts

Linear representation

Datasets

TriviaQA [eval], Winobias [eval]

Methods

Probing classifiers / MLP probing classifiers / Q16 classifier [primary]

Related findings

IC-181, IC-182, IC-184

Extraction

automatic-extraction