anchor
Decode intermediate activations through the model's output head to read what the model would predict at that depth.
Findings
- IC-019Visual token representations in LLaVA-1.5 evolve to align with interpretable text tokens [primary]
- IC-1015GPT-J and 10 other LLMs exhibit overthinking: calibrated accuracy given incorrect few-shot demonstrations peaks at a critical layer then declines, and ablating 5 false induction heads in late layers reduces the accuracy gap by 38.9% on average [primary]
- IC-106Logit lens on LLaVA and InstructBLIP image representations shows higher internal confidence for objects present in the image than for hallucinated objects [primary]
- IC-107Linear orthogonalization of LLaVA and InstructBLIP image features against text embeddings removes hallucinated objects at 83-86% individual rate versus 7-16% for correctly detected objects [supporting]
- IC-108Per-patch logit lens confidence in LLaVA localizes objects spatially, achieving mAP 79.90 on ImageNet segmentation, 8.03% above raw VLM attention [primary]
- IC-148Language models represent semantically equivalent inputs from different data types (languages, code, images, audio) close together in intermediate layers, with the shared space scaffolded by the model's dominant language [primary]
- IC-1553GPT-J, GPT-2-XL, and Llama-13B decode approximately 48% of tested relations via a linear transformation on the subject representation, and this structure causally influences predictions [compared-to]
- IC-1555GPT-J's internal representations contain correct factual knowledge even when the model outputs falsehoods under repetition or instruction distraction prompts [compared-to]
- IC-218LLaMA3-8B and other LLMs solve arithmetic via a bag of independent heuristic neurons in middle and late MLP layers rather than a robust algorithm [supporting]
- IC-219The bag-of-heuristics mechanism in LLaMA3-8B fails on certain arithmetic prompts due to insufficient total logit contribution from heuristic neurons, not due to a lack of associated heuristics [primary]
- IC-547Llama3.2 3B and Llama3.1 8B exhibit saturation events in which the top prediction, once it appears at a given layer, remains unchanged through all subsequent layers [primary]
- IC-610In LLaMA2-7B-Chat, RAG hallucinations are causally driven by copying heads losing external context information during generation and by knowledge FFNs in mid-to-upper layers over-adding parametric knowledge to the residual stream [primary]
- IC-715Factual information deleted from GPT-J, LLaMA-2, and GPT-2-XL via ROME or MEMIT remains linearly recoverable from intermediate hidden states, with up to 89% extraction success at budget b=20 [builds-on]
- IC-715Factual information deleted from GPT-J, LLaMA-2, and GPT-2-XL via ROME or MEMIT remains linearly recoverable from intermediate hidden states, with up to 89% extraction success at budget b=20 [primary]
- IC-744The decoded vocabulary of a function vector often reflects the task's output space, but reconstructing a vector that matches this vocabulary distribution does not recover the FV's full causal effect. [primary]