Why do Llama-3.

Leo91 Intermediate 2d ago 339 views 12 likes 2 min read

I've been digging into why linear probes that detect "lying" or deception in LLM activations usually crumble the moment you move them to a new dataset. If you train a probe on one set of deceptive prompts and test it on another, the accuracy usually tanks because the probe latches onto "surface features" (basically dataset-specific quirks) rather than the actual concept of deception.

I was looking at the findings from arXiv:2609.02893v1, and it turns out this is mostly a subspace selection problem. The researchers tested Llama-3.1-8B-Instruct across three held-out deception datasets, and the gap between a standard probe and an "oracle" (a probe trained directly on the test set) is huge. But here is the interesting part: if you project the inputs onto a small subset of principal components (PCs) from the training distribution, you can actually get cross-domain transfer that almost matches the oracle performance.

The breakdown of the "Surface Feature" trap

The problem is that the weights a source probe relies on most heavily are often just noise or specific patterns found in the training data. To fix this, they used an LLM judge to score individual PCs. The judge looks at the examples that most and least activate a specific PC to see if they actually imply a "deception direction" that makes sense across different domains.

By filtering for these high-scoring PCs and probing only on that subspace, they managed to close the baseline-to-oracle gap significantly:

  • Insider Trading Report dataset: gap closed by 78%
  • Sandbagging dataset: gap closed by 25%

This suggests that the "truth" or "deception" signal is actually there in the activations, but it's buried under a layer of distribution-specific garbage. The directions that actually transfer are more abstract and can be described in natural language, whereas the ones that fail are just overfitting to the source dataset's style.

Practical take on implementing this AI workflow

If you're trying to build a robust deception detector using probes, don't just fit a linear model to your activations. You should probably be looking at the PCA of your activation space first.

1. Extract activations for your training set.
2. Run PCA to find the principal components.
3. Use an LLM to analyze the "extreme" activations for each PC to identify which ones actually represent the concept you're looking for.
4. Project your test data onto only those specific PCs before running the probe.

This feels like a much more reliable way to handle OOD (out-of-distribution) data than just hoping your training set is diverse enough. It's basically a manual feature selection process using an LLM as a semantic filter.

For anyone trying to replicate this, you'll need the Llama-3.1-8B-Instruct weights and a way to hook into the internal activations (like using TransformerLens or similar). The key is that the "transferable" directions are the ones that an LLM can actually describe in plain English, while the "overfit" directions are just mathematical artifacts of the training set.

Help Wanted

All Replies (4)

T
TaylorDreamer Intermediate 2d ago
Probably happens because the "truth" direction shifts depending on the specific topic being discussed.
0 Reply
R
Riley2 Advanced 2d ago
Do you think using a contrastive approach during training would help stabilize the vector across datasets?
0 Reply
D
DrewCrafter Novice 2d ago
Maybe, but wouldn't that risk collapsing the latent space if the negative samples aren't diverse enough?
0 Reply
J
JordanGeek Expert 2d ago
ran into this with a gpt-4 experiment, the vectors just dont generalize well across domains.
0 Reply

Write a Reply

Markdown supported