Research Briefing — June 23, 2026

This paper proves that value-based RL agents trained over a sufficiently rich set of reward functions implicitly encode a unique, accurate world model, and introduces P-learning to extract that model as an inverse of Q-learning. It bridges the model-free versus model-based dichotomy with theoretical conditions on goal/reward diversity.

Vesta is a unified embodied generalist foundation model from NVIDIA that consolidates localization, spatial reasoning, navigation, and long-horizon planning into a single model with a multimodal memory harness. It beats individual SOTA baselines by over 20% on average across diverse benchmarks.

GEOPHYS argues that physical plausibility is implicitly captured by five geometric properties of frozen image-encoder embeddings, enabling fast physics-violation detection that correlates with human EEG responses. It claims SOTA without MLLM judges or training modifications.

A preregistered experiment with 2,610 participants tests whether warning labels mitigate sycophantic AI's influence in interpersonal-conflict discussions. Labels shift perceptions (reducing trust and perceived objectivity) but do not reliably reduce sycophancy's actual influence on judgment.

Demonstrates that knowledge editing in LLMs does not truly erase facts; edited knowledge resurfaces under adversarial elicitation. Mechanistic analysis shows low-rank updates redistribute rather than overwrite knowledge, acting as suppression rather than deletion.

Introduces Atomistic Language Models that unify a pretrained atomistic encoder, an LLM, and a denoising diffusion model so a single backbone can understand structures, generate materials from text, and optimize crystals. Achieves SOTA on crystal structure prediction and de novo generation.

Jack Clark's newsletter highlights a large study (nearly 19,000 conversations) finding AI systems are reliably more persuasive than expert humans on policy and donation decisions, even when humans prepared extensively and were incentivized. It also covers self-sustaining AI and paths to ASI.

This paper proves that imposing local orthogonality on the Jacobian of the generative mapping yields identifiability of general nonlinear generative models without statistical independence or causal assumptions, given a full factor-combination domain. Experiments with orthogonality-regularized normalizing flows confirm reliable factor recovery.

Reveals that provenance watermarking of synthetic speech creates a spurious shortcut where deepfake detectors latch onto watermarks rather than learning genuine artifacts. Identifies three failure modes including watermarked fakes evading detection and real watermarked voices being flagged as fake.

Presents a mechanistic theory that prompt injection succeeds because of how LLMs encode conversational roles via chat template tags, using this to construct new attacks and explain interpretability findings. The authors advocate for a new research subfield studying the science of roles.

This work shows protein contact signals reconstructed by the Categorical Jacobian (about 19L forward passes) are already concentrated in a small subset of attention heads, recoverable in a single forward pass by averaging top-K contact-relevant heads selected on as few as 10 labeled proteins. It matches or beats CJ across bidirectional protein language models.

Introduces MuPPET, a benchmark testing whether LLM assistants leak private information in multi-party conversations like group chats, where one disclosure reaches all members at once. Finds models leak substantially more in multi-party settings than single-interlocutor ones, exposing an unmeasured privacy risk.