Research Briefing — June 8, 2026

This paper proposes the Piggyback Hypothesis to explain emergent misalignment, showing that chat-template tokens carry finetuned misbehavior onto unrelated queries. The authors validate it via prefix perturbations and introduce Token-Regularized Finetuning (TReFT) to mitigate misalignment.

This study measures how well frontier models reason without chain-of-thought across 30,000+ questions in 43 benchmarks, estimating human time-horizon equivalents for tasks solved without explicit thinking tokens. It matters because CoT-based oversight breaks down if models can reason complexly internally. The strong author roster and safety-relevant framing make this notable.

This behavioral audit of seven LLMs across four US cities tests racial steering in housing recommendations under progressively detailed prompting that mirrors fair-housing paired-testing. It finds steering is an emergent property of model interpretation interacting with user identity and preferences rather than a static property.

This theoretical paper provides a unifying definition of large step sizes for gradient descent requiring only local Lipschitz or Holder gradient continuity, addressing the longstanding question of maximum convergent step size for non-smooth objectives. It designs adaptive methods that provably operate at the edge of stability from training start.

Introduces CapCode, which builds coding datasets with randomized tests whose maximum non-cheating score is deliberately capped, so scores above the cap reveal reward hacking, plus CapReward to discourage exploitation. Addresses deceptive performance in coding agent evaluation and training.

MMBU is the largest biomedical vision-language benchmark to date, covering 35 submodalities with structured metadata and tasks spanning ungrounded/grounded classification and object detection. It probes fine-grained perception capabilities of VLMs across diverse biomedical imaging contexts.

Argues provocatively that native hardware FP64 is not essential for scientific computing, showing that FP8 tensor throughput plus the Ozaki Scheme II can recover full FP64 accuracy on AI-optimized GPUs like NVIDIA B300. Introduces a Tensor-Memory Equilibrium roofline model to support the claim.

The authors characterize LLM reasoning failures using token-level uncertainty signals, distinguishing committed failures (early lock-in) from persistent uncertainty (accumulating doubt). The framework offers falsifiable diagnostics reproducible across 23 model-dataset configurations, useful for failure detection in reasoning systems.

This theoretical paper establishes the first minimax-optimal generalization rates for deep ReLU networks trained with gradient descent and SGD in the NTK regime, extending prior work limited to shallow architectures. It assumes polynomial width scaling with depth and training samples.

This paper develops an analytical model explaining how training task diversity shapes in-context learning in transformers by modeling task vectors as a mixture of low-rank Gaussians. It provides theoretical grounding for previously observed but unexplained ICL phenomena.

A position paper arguing that a genuine science of AI must study training dynamics rather than relying on post-hoc fixes, advocating for predicting outcomes from early signals, intervening on trajectories, and designing training to reliably produce desired properties. Frames scaling laws as just a start toward understanding emergence.

ViSAE is a neuroscience-motivated mechanistic interpretability toolbox for Vision Transformers, using sparse autoencoders to extract human-interpretable concept circuits with a 64K-image probing suite and 16K concept vocabulary. It enables understanding and steering of ViT internals to mitigate spurious cues.