Also Worth Noting - 2026-05-27

Also Worth Noting

02 [Training] Self-Improving Language Models with Bidirectional Evolutionary Search Standard self-improvement loops stall because best-of-N sampling only explores where the model already assigns high probability. Bidirectional Evolutionary Search couples forward candidate generation with backward search from a target, generating training samples that sit outside the model's current probability mass without requiring a dense reward model. The sparse verification signal problem, which is the actual bottleneck in self-improvement pipelines, gets bypassed rather than solved directly. Teams running post-training self-improvement loops should evaluate BES before investing in denser reward modeling infrastructure. link

03 [Inference] D2-Monitor: Dynamic Safety Monitoring for Diffusion LLMs via Hesitation-Aware Routing Diffusion LLMs expose intermediate denoising states that carry safety-relevant signals invisible to any single-pass content filter, meaning teams deploying a diffusion LLM today are running it without the equivalent of a content monitor. D2-Monitor identifies trajectory-level hesitation signals that predict when lightweight probes are likely to fail, then routes those cases to stronger inspection. The architecture is designed for always-on production use, not offline auditing. Any team evaluating D-LLMs for deployment should treat safety monitoring as an open infrastructure gap, not a solved one. link

04 [Agent] MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research Screenshot-diff heuristics and human raters have been the de facto judging methods for mobile GUI agents, making RL training signals expensive and inconsistent. MobileGym replaces that with deterministic JSON-state judging: the full environment state is captured, forked, and compared as structured JSON, so outcome verification is cheap and reproducible. A single server hosts roughly 400 MB per instance, enabling hundreds of parallel rollouts at a cost that makes online RL practical. Teams doing mobile agent research can now run verifiable RL experiments at a scale that was previously out of reach. link

05 [Training] Understanding Data Temporality Impact on Large Language Models Pre-training The standard practice of shuffling pre-training corpora actively degrades a model's ability to place facts in time, and the field has mostly ignored this. Training on chronologically ordered data improves temporal grounding, measured against a benchmark of over 7,000 temporally grounded questions designed to test whether models correctly associate facts with the right time periods. Corpus shuffling has no well-documented accuracy benefit that offsets this cost. Teams curating pre-training data pipelines should treat temporal ordering as a variable worth controlling, not a default to accept. link

06 [Application] Efficient and Scalable Provenance Tracking for LLM-Generated Code Snippets Winnowing-based fingerprinting catches verbatim training-data reproduction in LLM-generated code, but its linear-time search makes it impractical against billion-scale corpora. This work adapts classical fingerprinting to scale to the output volumes of production code LLMs without requiring access to model internals. The result is a concrete detection path for license compliance that legal and engineering teams can run on generated code as it ships. Teams using LLM code generation in production have a deployable provenance-tracking option that does not depend on model cooperation. link