Also Worth Noting - 2026-07-19

Also Worth Noting

02 [Inference] Kernelized Linear Attention: Breaking the Capacity Wall with Symmetric Cones Linear attention's recall failure is a geometry problem, not a tuning problem. KATA reframes associative recall as spherical packing, then derives feature maps from first principles by certifying nonnegative attention weights through a self-dual homogeneous cone. Rank-one positive semi-definite features yield a favorable capacity-interference tradeoff and recover a parameter-free convex output gate. Teams chasing constant-cost inference without sacrificing recall now have a principled path rather than another empirical patch. link

03 [Training] A Systematic Evaluation of Trajectory Data Curation for LoRA Fine-Tuning of Code Agents Past a quality threshold, adding more low-quality trajectories actively hurts LoRA fine-tuning performance. Working across 67,074 trajectories from SWE-trajectory with Qwen2.5-Coder-7B-Instruct, the study evaluates a two-axis Efficiency-and-Style scoring framework across 16 controlled conditions. Quality and quantity interact non-linearly, which breaks the common SFT assumption that more data is always better. Teams fine-tuning open-weight code agents should filter aggressively before scaling trajectory volume. link

04 [Theory] Quantifying Diversity of Thought: A Predictive Law of Weighted LLM Ensemble Lift Ensemble lift can be calculated before running inference, not just measured after. Decomposing lift into rescue mass and damage mass yields a compact heuristic built around accuracy-adjusted correctness correlation, the accuracy gap, and collective accuracy of any model pair. The law was verified across 767,520 inferences from ten open-weight models. That turns ensemble selection from trial-and-error into a pre-compute check, saving substantial inference budget. link

05 [Agent] WAR: Workload-Aware Rollouts for Synchronous Agentic Reinforcement Learning Long-horizon trajectories stretching to tens of thousands of tokens create severe load imbalance across workers, and that imbalance is now the dominant wall-clock bottleneck in synchronous agentic RL. WAR addresses this by jointly optimizing decoding and scheduling based on runtime load: model-free speculative decoding under low load, and rebalanced scheduling under high load. The RL algorithm itself stays unchanged. Teams can adopt WAR as a drop-in systems change without touching their training logic. link

06 [RAG] Evidence Interfaces Shape How Retrieval-Augmented Readers Use Support A top-k RAG accuracy score conflates two separate failures: the retrieval window drops part of the support chain, or the reader does not use the evidence form it receives. A reader trained on raw context performs differently than one trained on structured evidence interfaces even when retrieval is identical. Testing across three support-annotated multi-hop QA benchmarks, the study shows that standard eval scores mask both problems simultaneously. Teams debugging RAG pipelines should separate retrieval-availability failures from reader-interface effects before drawing any conclusions about where the pipeline breaks. link