• NeoHorse-1 Turns Routing Logs Into a Training Flywheel. Its 4B model’s macro average across 11 benchmarks rose from 58.94 to 64.87, offering a measurable path from production telemetry to capability gains.

• AuK Unifies Speech Generation and Editing Behind One Interface. About 3.03 billion instruction-audio examples support five task categories, while AuK-Flash achieves a 4.5× end-to-end speedup with four inference steps.

• When Long-Horizon Agents Cannot Learn, Improve the Environment’s Feedback First. Feedback-enriched environments produced consistent gains across model sizes and RL algorithms on SciWorld and BFCL.

• Miles Breaks the RL Training Loop Into Verifiable Components. Customizable training components aid system changes and debugging, though one large-scale case cannot establish broad performance or cost advantages.

Also Notable

• VLA Models Keep Training Their World Models at Test Time as Environment Dynamics Shift. — WorldAgen jointly predicts states and actions, exploring continuous adaptation to unfamiliar environments..
• A Biomedical Benchmark Tests Whether Dense Retrieval Can Satisfy Conceptual, Relational, and Compositional Phenotype Constraints. — OntologyBench contains 471,854 training pairs and 125,744 evaluation pairs..
• More User History Does Not Always Improve Personalization. — The study learns minimal sufficient profiles that support personalized generation with less context and redundant information..
• Repository Retrieval Is Shifting From Isolated Snippets to File-Centered Navigation. — RepoNav preserves structural relationships for repository tasks that require locating a small set of relevant files and functions..
• Attention Weights May Not Faithfully Explain Instance Contributions in Digital Pathology. — CAR-MIL uses counterfactual regularization to improve attention explanations when fine-grained labels are expensive..
• Do Reviewers Still Reward Complex Language When Polished Prose Becomes Cheap? — An analysis of 124,000 ICLR reviews tracks how reviewer preferences change across years..
• PlannerForge Connects Autonomous-Driving Scenario Creation, Retrieval, Editing, Execution, and Analysis. — Its LLM agent targets fragmented safety-testing workflows for motion planners..
• Implicit Neural Image Representations Need Fast Encoding and Sub-Millisecond Decoding to Become Practical. — PIC reassesses the efficiency and deployment potential of INR-based image coding..
• RoPE May Not Cause Attention Sinks After All. — The study points instead to self-concentration and value non-mixing, linking them to activation outliers in low-bit quantization..
• Multilingual Bias Research Is Tracing How Stereotypes Reach Model Answers. — Probes, attribution patching, and sparse autoencoders track the path from internal representations to outputs across languages..

Read the full edition →