Research Briefing — June 18, 2026

Evaluates LLM lie detectors by building belief-verified model organisms that demonstrably hold a belief contrary to what they state, plus a prompted-lying testbed. Finds that activation- and logprob-based detectors scale positively when lying is prompted but drop sharply when lying is trained in, casting doubt on current detectors' reliability.

Rift investigates whether deception leaves an internal signature in LLMs by contrasting a sleeper agent that lies on trigger against a naive liar producing identical wrong answers. It finds deceptive passes carry a conflict signature with higher residual rank, enabling lie identification with 100% accuracy and no labels.

Shows that interventions on sparse autoencoder features intended to suppress unsafe model behavior are unreliable, because the behavior can be recovered via residual-space optimization even after clamping. This challenges latent-space safety defenses that assume SAE features are actionable handles for control.

DO AS I DO reconstructs and retargets monocular RGB human videos to multi-fingered dexterous robot hands, recovering hand-object interactions from in-the-wild egocentric and exocentric sources. It aims to unlock abundant RGB-only human video as scalable dexterous manipulation data.

This work shows that an LLM's KV cache acts like a notebook of memoized conclusions, where a field's own key/value vectors drive under 1% of downstream decisions. This enables editable and composable caches, allowing field corrections via chain-of-thought at roughly 1% compute.

STATEWITNESS is an activation explainer for deception auditing that uses a separate decoder to read a target model's hidden states and produce natural-language reports about suspicious behavior. It reaches 0.916 mean AUROC across seven deception datasets, outperforming black-box and probe baselines.

Reversal Q-Learning is an off-policy RL algorithm that trains a flow policy on prior data using an expanded MDP treating flow refinement steps as actions, generating virtual on-policy trajectories by reversing flows. It applies bias-and-variance reduction to mitigate the curse of horizon.

Self-CTRL trains language models so their self-explanations and actual behavior remain consistent, by reinforcement learning that updates either explanations or behavior. Tested on probabilistic reasoning tasks, it improves the correlation between self-reported and measured biases, supporting auditability and trust.

SC3-Eval adapts a pretrained video foundation model into an accurate robot policy evaluator by enforcing forward-inverse dynamics, multi-view, and out-of-distribution consistency. It offers a scalable alternative to expensive real-world policy evaluation.

Extends UK AISI work on prefill awareness, showing several frontier models can distinguish tampered assistant-side content from genuine content even under conservative elicitation and in low-stakes settings. This matters because prefill-based methods underpin many safety evaluations (jailbreaking, persona, introspection), and awareness could confound pre-deployment safety claims.