Research Briefing — June 10, 2026

This paper shows that AI-assisted peer review systems can be manipulated through superficial rephrasing of manuscript abstracts, improving review outcomes without changing scientific content. The vulnerability holds across disciplines and even without knowledge of the reviewing model, raising concerns about the integrity of AI-mediated scientific evaluation.

This work shows that one-shot GRPO training on a single biased example can induce systematic, generalizing bias in aligned LLMs, revealing that alignment guardrails can be overridden trivially. It exposes a critical vulnerability in RL-based post-training.

Documents that low-bit KV cache quantization can silently destroy LLM safety alignment, with Mistral-7B losing 15 percent of refusals at negligible perplexity change and no universally safe bit-width existing. It traces the cause to safety features occupying a low-dimensional activation subspace and proposes mitigation.

This trustworthiness audit examines whether converting instruction-tuned LLMs into reasoning models preserves alignment behaviors like safe refusal and bias avoidance. It finds that SFT, RL, and distillation-based reasoning conversions are not behavior-preserving across six trustworthiness dimensions.

Kwai Keye-VL-2.0 is an open-source 30B Mixture-of-Experts multimodal foundation model for long-video understanding and agentic intelligence, notably the first to adapt DeepSeek Sparse Attention to GQA-based multimodal architectures for lossless 256K context processing. It includes optimized training and inference infrastructure for hour-level video.

Presents GitInject, an open-source framework that evaluates prompt injection vulnerabilities in real live GitHub CI/CD workflows by provisioning ephemeral repositories and triggering actual workflow runs, rather than simulating tool calls. It exposes supply-chain risks from agents ingesting untrusted content with elevated permissions.

Introduces the first computational framework that directly tackles the dual problem for revenue-optimal multi-item multi-bidder auctions, parametrizing Lagrange multipliers with neural networks to generate certified revenue upper bounds under dominant-strategy incentive compatibility. It addresses a fundamental open problem in mechanism design.

Studies whether simple test-time gradient guidance of flow/diffusion policies can improve RL performance while leaving stable supervised training intact, avoiding backpropagation through denoising. From a strong Berkeley robotics-RL group including Sergey Levine.

This paper introduces the Ito map, an any-step stochastic flow map that predicts future states from an intermediate state and Brownian path in a single pass, enabling exact distillation for stochastic dynamics. It provides differentiable access to posterior samples for inference-time control and steering.

Introduces CIAware-Bench to measure whether frontier models can detect when their trajectories have been modified by AI control interventions, across four task domains while varying watermarking and side-tasks. Probes a risk that models may infer monitoring and adapt to evade control.

Introduces PreActBench, a 1000-example benchmark for Predictive Monitoring, the task of inferring whether a partial agent action trajectory will culminate in an unethical action before it is executed. It shifts safety evaluation from retrospective harm detection to proactive prevention.

Introduces VFUSE, a mechanistic interpretability approach that trains sparse autoencoders on protein diffusion-transformer activations to audit models like RoseTTAFold3 and RFDiffusion3 for hazard-aware features. Linear probes in the SAE latent space detect hazardous protein designs better than in the original representations.