Glonce

New method spots false claims by reading LLM activations

researchThe Glonce editorsarXiv🔥72Glonce rating72Impact65Novelty70Relevance100Surprise50Players32Credibility80

Most misinformation detectors either look at surface-level linguistic features or retrieve outside evidence to check a claim. A new paper takes a different route: it treats truthfulness as a geometric property of a language model's internal representation space. The authors build a detection framework around activation engineering, first eliciting a 'misinformation direction' in the model's residual stream by contrasting activations from paired truthful and false statements, following the difference-in-means principle behind Contrastive Activation Addition (CAA). At inference time, the last-token activation of an unseen claim is projected onto this direction, and that projection is fed into a Multilayer Perceptron (MLP) for classification. The procedure needs no fine-tuning of the underlying model, no external evidence retrieval, and no task-specific supervision beyond the contrastive pairs used to estimate the direction in the first place.

The authors test the approach on 11 models from the Gemma, Llama, and Qwen families, ranging from 270M to 12B parameters, across three fact-checking benchmarks: AVeriTeC, LIAR, and FACTors. They report that the falsehood direction is recoverable across model scales and architectural families, and that last-token projection matches or surpasses zero-shot and few-shot prompting baselines on LIAR and FACTors, with the largest gains showing up in smaller models. Performance on AVeriTeC is weaker, which the authors attribute to its evidence-grounded labeling scheme, where a claim's truth depends on evidence the model was never shown rather than patterns it already encodes.

The authors conclude that truthfulness is a structured, linearly separable concept in the latent space of pretrained language models, and frame interpretability-driven detection like this as a practical complement to retrieval-based fact-checking pipelines rather than a replacement for them. Code is available on GitHub at Malta-Lab/LaFaCt.

Key facts

  • The framework locates a 'misinformation direction' in a model's residual stream via difference-in-means contrast between true and false statement activations (the CAA approach), needing no fine-tuning or external evidence retrieval.
  • Classification works by projecting the last-token activation of a new claim onto this direction and feeding it to an MLP classifier.
  • Tested on 11 models (270M to 12B parameters) from the Gemma, Llama, and Qwen families across three benchmarks: AVeriTeC, LIAR, and FACTors.
  • Last-token projection matches or beats zero-shot and few-shot prompting baselines on LIAR and FACTors, with the largest gains on smaller models; AVeriTeC results are weaker, which the authors attribute to its evidence-grounded labeling scheme.
  • Code is released on GitHub under Malta-Lab/LaFaCt.

Why it matters

Most misinformation detectors either scan for surface linguistic cues or need to retrieve outside evidence to verify a claim. This work argues truthfulness has a geometric signature inside a model's own activations: a linearly separable direction that can be extracted without touching the model's weights. If that holds up, it opens a lightweight path to fact-checking that skips an external knowledge base or fine-tuning, needing only a small set of contrastive true/false examples to locate the direction once.

Who it affects

Teams building fact-checking or content-moderation tooling on top of open-weight models, and researchers studying how LLMs internally represent truth versus falsehood. The method was tested on Gemma, Llama, and Qwen models spanning 270M to 12B parameters, so it targets open-weight, inspectable models rather than closed API-only systems.

How to use it

The authors release code on GitHub at Malta-Lab/LaFaCt. Building a detector means running paired true and false statements through the target model to compute the activation difference (the CAA-style step), then training a small MLP classifier on the resulting projected last-token activations. No fine-tuning of the underlying model and no external evidence-retrieval pipeline is required beyond the initial contrastive pairs.

How solid is it

The method was evaluated across 11 models from three families (Gemma, Llama, Qwen) and three fact-checking benchmarks (AVeriTeC, LIAR, FACTors), a reasonably broad sweep for one paper. Last-token projection matched or beat zero-shot and few-shot prompting baselines on LIAR and FACTors, with the biggest gains on smaller models. The text does not give specific accuracy or F1 figures for the individual benchmarks or model families, and does not describe how many contrastive pairs were used or the MLP classifier's architecture, so the exact size of the improvement and the classifier's details cannot be checked from the paper's text alone.

Risks and caveats

Performance dropped on AVeriTeC, which the authors attribute to its evidence-grounded labeling scheme, suggesting the method may struggle with claims whose truth depends on external, up-to-date evidence rather than patterns already encoded in the model. Because the approach relies entirely on a model's pretrained internal representations, it is unclear how it would generalize to genuinely novel misinformation, adversarially crafted claims, or models whose training data or architecture differ substantially from the eleven tested here.

“The falsehood direction is recoverable across model scales and architectural families.”

— the paper's authors