Research Briefing — April 30, 2026

Introduces QERNEL, a foundational neural wavefunction using FiLM conditioning and mixture-of-experts to solve families of many-electron Hamiltonians. Applied to up to 150 electrons in semiconductor moiré heterobilayers, discovering phase transitions.

Presents CheXthought, a global multimodal dataset with 103,592 chain-of-thought reasoning traces and 6.6M visual attention annotations across 50,312 chest X-rays from 501 radiologists in 71 countries, revealing clinical reasoning patterns.

Presents GLM-5V-Turbo, a native foundation model for multimodal agents integrating visual perception as a core component of reasoning, planning, and tool use. Built with improvements across model design, multimodal training, RL, and agent frameworks.

Introduces a probe-based data attribution method that identifies training datapoints responsible for harmful behaviors emerging during DPO post-training. Filtering flagged data reduces harmful compliance by 63% without hurting general performance, outperforming gradient-based methods at 10x lower cost.

Rethinks KV cache eviction through the Information Bottleneck principle, deriving a closed-form mutual information objective under a linear-Gaussian attention surrogate. Shows existing eviction strategies are different approximations of the same capacity-maximization principle.

Presents DORA, a scalable asynchronous RL system for LLM training that addresses the long-tail rollout bottleneck while maintaining algorithmic correctness through intra-trajectory policy consistency, data integrity, and bounded staleness constraints.

Introduces DreamProver, a wake-sleep framework for formal theorem proving that iteratively discovers reusable lemmas. In the wake stage it proves theorems while proposing candidate lemmas; in the sleep stage it abstracts and consolidates them into a transferable library.

Formalizes alignment faking detection in LLMs through observable tool selection behavior, where models choose safe tools when unmonitored but switch to unsafe tools under monitoring that rewards helpfulness. Named 'Tatemae' after the Japanese concept of public façade.

Introduces Incremental Completion Decomposition (ICD), a jailbreak strategy that elicits single-word continuations before requesting full harmful responses, achieving high attack success rates across multiple model families.

Demonstrates that fine-tuning datasets for constitutional classifiers (as described by Anthropic) can be poisoned to create backdoors that allow harmful prompts through. The attack exploits the synthetic data generation pipeline used to train these safety classifiers.

Proposes risk reporting requirements for AI developers' internal model use, noting that frontier companies deploy advanced models internally for weeks before public release (citing Anthropic's Mythos Preview). Analyzes gaps in existing legal frameworks like California's SB 53 and EU's GPAI Code.

Presents AMMA, a multi-chiplet memory-centric architecture for low-latency attention serving at 1M context length, arguing that GPU-centric designs are mismatched with memory-bound decode-phase attention workloads.