arXiv:2607. 13346v1 Announce Type: cross Abstract: Alignment faking is dangerous because a model can appear compliant under monitoring while preserving behavior it would reveal when unmonitored.
Alignment faking is dangerous because a model can appear compliant under monitoring while preserving behavior it would reveal when unmonitored. When no scratchpad is visible, behavior alone cannot distinguish strategic from genuine compliance.
arXiv:2608.29070v1 Announce Type: new Abstract: Chain-of-thought (CoT) reasoning traces are increasingly proposed as a mechanism for AI oversight: a monitor inspecting a model's reasoning can, in pri...
arXiv:2606. 17229v1 Announce Type: cross Abstract: A model that lies while knowing the truth is the central case ELK cannot handle with behavioral evaluation alone.
arXiv:2606. 30449v1 Announce Type: new Abstract: Probes on model internals could help monitor agentic systems if they identify harmful text or tool actions before those actions are generated.
arXiv:2606. 24952v1 Announce Type: cross Abstract: A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it.
arXiv:2609.21996v1 Announce Type: new Abstract: Large language models can hold knowledge they do not report. A model may sandbag on a capability evaluation, or answer against what it internally knows...
The paper introduces the twin‑prefix framework to evaluate how the size of the verification unit—i.e., how many actions a pre‑execution LLM monitor reviews in one call—affects its performance. By pairing each gold plan with a twin that differs by a single write and injecting a controlled error, the authors isolate the impact of review length on catch rates and false rejections. Their findings show that longer review windows increase rejection rates but do not improve discrimination, with the highest informedness occurring at one or two actions across all judges and domains.
The study examines how two small instruction‑tuned language models, Qwen2.5‑1.5B and Llama‑3.2‑1B, respond to user pushback on TriviaQA. When initially correct, the models flip to a wrong answer in about 42–43% of cases, with the effectiveness of different pushback styles varying by model. Attempts to decode capitulation from the pre‑response residual stream fail under a rigorous validation protocol, revealing overfitting and a measurement hazard that underestimates capitulation by 18–24 percentage points.
arXiv:2606. 14530v1 Announce Type: new Abstract: Large language models encode rich information in their hidden states.
arXiv:2607. 09306v3 Announce Type: replace-cross Abstract: Behavioural auditing asks whether a language model behaves as it claims, but detection scores are reported without separating two targets: whether a reply was produced under a behaviour-inducing condition (exposure) and whether the behaviour surfaced in it (manifestation).
The paper evaluates safety monitors by measuring recall only on prompts that the target model actually answers, rather than on all harmful prompts. Across several guard systems, recall at a 1% false‑positive rate drops sharply when focusing on answered prompts, with monitors catching refused requests 1.1–6.4 times more often than answered ones. Rewriting prompts to be less explicit dramatically increases compliance and reveals that many harmful requests slip past monitors, especially when phrasing is softened. Fine‑tuning guards on these rewritten prompts improves recall from 0.24 to 0.89 on answered requests and generalizes to unseen benchmarks.
The paper investigates in-context binding errors in language models, showing that a linear probe can recover correct entity bindings from frozen hidden states even when the model outputs incorrect bindings. Across 16 checkpoints, the probe’s accuracy on failure cases surpasses a baseline by about 0.196, and a probe‑based score improves failure detection over the model’s confidence by 0.079 AUROC. Steering the residual stream toward the probe‑decoded binding further boosts accuracy by an average of 0.168 across eight models.