Changing sparse autoencoder (SAE) codes can alter both intervention direction and magnitude, so comparing intervention effects requires a strength control specifying what is held equal. Norm matching...
The paper introduces AMRA, a weight‑editing technique that mitigates abliteration—an attack that removes refusal capabilities from large language models by projecting weight matrices orthogonal to a refusal direction. AMRA obscures the refusal signal through rank‑$k$ updates to residual stream writer matrices, replaces refusal‑inducing activations with random aliases, and adjusts downstream reader matrices to maintain original behavior. Experiments on Llama‑3‑8B and Gemma‑2‑9B show significant improvements in post‑abliteration refusal scores with minimal impact on overall model performance.
The study investigates whether open‑weight language models can introspect on their own internal computations. Using the Open‑Weight Masked Introspection (OWMI) framework, researchers intervened on various internal components of eight models and asked them to report whether changes had occurred. Across 78,000 measurements, none of the models reliably distinguished real interventions from sham ones, with AUROC values essentially at chance. Why It Matters: The findings suggest that current open‑weight models lack the ability to audit their own internal states, highlighting a limitation for oversight that relies on a model’s self‑reporting.
arXiv:2608. 02830v1 Announce Type: cross Abstract: Many-shot in-context learning (ICL) lets vision-language models (VLMs) adapt from image--label demonstrations without weight updates, and is widely assumed to improve as more demonstrations are supplied.
arXiv:2607. 14111v1 Announce Type: cross Abstract: Can small language models detect and report on perturbations their own internal activations?
The paper introduces Propose-Then-Project Unlearning (PTP-U), a method for large language model unlearning that first weakens target knowledge associations and then projects non-target outputs back to the original distribution to recover capabilities. PTP-U aims to satisfy forgetting constraints while preserving fluent generation and performance on non-target tasks. Experiments on three benchmarks show PTP-U achieves a strong forgetting‑retention trade‑off, with 81.22%–91.03% forgetting and an average of 94.20% non‑target utility.
arXiv:2606. 08365v1 Announce Type: cross Abstract: Sparse autoencoder (SAE) features are increasingly used to steer language models, but feature steering is rarely clean: the same intervention can behave inconsistently across contexts and perturb unrelated features.
Large language model unlearning aims to suppress designated undesirable knowledge while preserving benign capabilities. Many unlearning objectives focus on suppressing undesired answers, while recent target-guided variants specify replacement behavior but still leave update locality largely unconstrained.
arXiv:2608. 04021v1 Announce Type: cross Abstract: Cloze-style probes that vary how often a target token appears implicitly assume that more copies of a target affect prediction the same way regardless of where the readout slot sits.
arXiv:2606. 10989v1 Announce Type: new Abstract: Large language model unlearning aims to suppress designated undesirable knowledge while preserving benign capabilities.
arXiv:2610.27724v1 Announce Type: new Abstract: Dense linear probes assess task readability from language model (LM) representations, but do not directly reveal how many features suffice for predicti...
FishBack introduces a pullback Fisher geometry approach for activation steering in transformers, challenging the common Euclidean assumption of intermediate activation spaces. By deriving a closed‑form steering direction based on the Fisher information metric of the softmax layer, the method achieves target concept changes with minimal off‑target distortion, especially in early and middle layers. Experiments on GPT‑2 Small, Llama‑3‑8B, and Qwen3‑8B demonstrate significant reductions in off‑target KL divergence compared to existing steering baselines.