truthfulness

This paper identifies surface-level feature leakage in truthfulness benchmarks like TruthfulQA, where models can cheat by exploiting answer form differences, and introduces Audit-Prune to clean benchmarks, ensuring more reliable evaluations.

This research paper proposes Provenance Density, an evidence-visualization interface to help users distinguish AI-generated text by showing verified claims, mitigating the Fluency Trap where users trust fluent but fabricated content.

Joseph Decker evaluates 16 AI models on truthfulness for his product Condensr, discovers that the leaderboard winner fabricated content five times in an audit, and instead ships the second-place model which had zero fabrications. The post details the evaluation process, a bug in the LLM judge that penalized accurate summaries due to truncated transcripts, and the importance of custom evals over generic benchmarks.

Introduces ReDe, a framework that denoises reasoning traces by filtering irrelevant and repetitive steps to improve hallucination detection in large reasoning models, achieving up to 87.32 AUROC on TruthfulQA.

GrAInS is a contrastive gradient-based method that uses Integrated Gradients to identify influential tokens and construct steering vectors for inference-time steering of LLMs and VLMs, improving truthfulness and reducing hallucinations without degrading fluency.

A new paper shows that scaling inference compute via methods like self-consistency improves LLM accuracy in math and code but fails to improve truthfulness in domains without external verifiers, as model errors are too correlated.

ConflictScore is a new metric that quantifies how well language models acknowledge conflicting evidence in their grounding documents, decomposing responses into atomic claims and measuring conflict balance. The paper also introduces ConflictBench, a benchmark covering diverse conflict forms, and shows the metric can improve truthfulness on TruthfulQA.

Elon Musk claims Grok is maximally truthful, referencing a claim that Fable 5 (hypothetical) lies 96% of the time.

TriEval is a new pipeline for evaluating LLMs across bias, toxicity, and truthfulness simultaneously, designed to be resource-efficient and run on standard laptops. It has been tested on Llama 3 8B, Mistral 7B, Gemma 2 9B, and Claude Haiku, and is released as open source.

This paper investigates whether open-source quantized LLMs encode a linearly separable truthfulness signal in their hidden states. Across three 7B-8B instruction-tuned models, a linear probe on a single mid-network layer achieves 0.904-1.000 AUROC on hallucination detection benchmarks, outperforming sampling-based methods.

A reflection on how many AI models prioritize sounding confident over being truthful, using Claude as an example of a model that seems more focused on internal consistency and logical honesty.

This paper identifies a phase transition in language model scaling where below a critical parameter count, reasoning and truthfulness are anticorrelated, but above it they cooperate. It provides diagnostics and interventions for improving alignment across model families.

FineSteer is a novel inference-time steering framework that decomposes steering into conditional steering and fine-grained vector synthesis stages, using Subspace-guided Conditional Steering (SCS) and Mixture-of-Steering-Experts (MoSE) mechanisms to improve safety and truthfulness while preserving model utility. Experiments show 7.6% improvement over state-of-the-art methods on TruthfulQA with minimal utility loss.

TruthfulQA is a benchmark of 817 questions across 38 categories designed to measure whether language models generate truthful answers. The study found that the best model achieved only 58% truthfulness compared to 94% for humans, and larger models were generally less truthful—suggesting scaling alone is insufficient for improving truthfulness.