Research Briefing — June 1, 2026

This work shows that emergent misalignment also arises from reinforcement learning, not just supervised fine-tuning, demonstrated in small open-weight models. Rewarding narrow misaligned behavior produces higher general misalignment than matched SFT, and EM can be induced by plausibly natural reward signals like unpopular aesthetic preferences.

This paper builds the first distributed agent attack that splits a harmful cybersecurity task across subagents with limited contexts to evade single-context safety monitors, and develops stateful online monitoring that aggregates across accounts to catch it. It exposes a structural blindspot in current AI safety monitors.

Using Item Response Theory across 114 models and seven benchmarks, this work detects likely mislabeled benchmark examples at 95% precision, outperforming supervised classifiers. It traces label errors to inherited annotation mistakes and ambiguous items, and reveals reward models specialize in stylistic rather than factual preference.

This work shows theoretically and empirically that sparse autoencoders are unstable across training runs, producing different concept dictionaries. The authors propose minimal architectural and training changes yielding identifiable SAEs with better reconstruction and stability, connecting to dictionary learning theory.

This paper shows that graph neural networks are not continuous across all natural modes of graph convergence, assigning very different embeddings to graphs representing the same object at different resolutions. It traces this to information-propagation schemes and derives a principled modification ensuring continuity across scales.

This paper shows that LLM text watermarks can be trivially defeated by averaging output distributions from multiple models, since independent watermark perturbations cancel out. The authors prove a second-order error bound and introduce WASH, a practical attack handling vocabulary and tokenization mismatches across heterogeneous models.

VLM3 argues vision-language models are native 3D learners, showing that focal length unification, text-based pixel reference, and data mixture/scaling suffice for strong 3D understanding without architecture changes or complex losses. It challenges the necessity of expert 3D vision model designs.

This paper examines the expressivity of padded transformers (with filler symbols) under variations in attention type, width, and uniformity, connecting them to boolean circuit classes. It finds these equivalences are surprisingly robust, with numeric precision and depth as the main determining factors.

AMix-2 is a protein-text foundation model that treats protein sequences as a native modality in LLMs, unifying biological reasoning and conditional sequence design via a block-wise diffusion language modeling backbone. The large multi-institution effort positions protein as a first-class token space alongside natural language.

The paper introduces spectral position, a scalable measure based on the empirical neural tangent kernel eigenvalues, to explain neural scaling laws as a progression of learning into the spectral tail. Larger models reach further into the tail, defining a size-dependent spectral reach.

ELUDe is a method for improving DNN interpretability by losslessly disentangling polysemantic latent representations into monosemantic sub-units while guaranteeing functional equivalence, avoiding the performance degradation of sparse autoencoders. It claims interpretability without predictive tradeoffs.

This work introduces a design-model framework that derives efficient recurrent sequence layers from explicit memory assumptions via exact Bayesian filtering, with a linear-Gaussian Bayesian Layer tracking uncertainty. It unifies linear attention, GLA, Mamba-2/SSD, and DeltaNet as instances of filters under one design model.