Also Worth Noting - 2026-07-02

Also Worth Noting

02 [Inference] ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving Two decode workers at identical load can produce very different latencies, because each step pays the cost of loading every distinct expert the batch activates. ELDR predicts which experts a request will activate during generation by building an expert signature from its prefill activations, then routes to workers that already hold those experts in cache. Offline balanced K-means clustering organizes workers around expert affinity, keeping routing decisions fast. Teams serving Mixtral or DeepSeek at scale can cut per-step decode latency without touching model weights or changing batch sizes. link

03 [Training] CausalMix: Data Mixture as Causal Inference for Language Model Training Proxy-model approaches to data mixture optimization break the moment the underlying data pool changes, because they model correlation, not causation. CausalMix reframes mixture weights as causal interventions, estimating the effect each domain exerts on downstream performance independently of what else is in the pool. That separation means new domains can be added or swapped without retraining the proxy from scratch, a cost that makes methods like DoReMi prohibitive at scale. Teams managing evolving pretraining corpora can update mixture weights incrementally rather than restarting the optimization pipeline. link

04 [Agent] AutoTrainess: Teaching Language Models to Improve Language Models Autonomously Post-training is not just a coding problem, and that distinction is what most autonomous pipelines miss. AutoTrainess exposes planning, data construction, training execution, checkpoint evaluation, and experiment state management as a unified set of agent-computer interfaces, letting a single LM agent run the full post-training loop without human checkpoints between iterations. The system shows measurable gains on coding and math benchmarks over baseline post-training runs. For teams iterating on fine-tuning pipelines, this is a concrete step toward reducing the human hours that currently sit between each training cycle. link

05 [Eval] PerceptionRubrics: Calibrating Multimodal Evaluation to Human Perception Models scoring near-ceiling on standard VQA benchmarks drop measurably when evaluated against atomic, image-specific criteria, which means saturation reflects scoring granularity, not actual perception capability. PerceptionRubrics pairs 1,038 information-dense images with over 10,000 instance-specific rubrics derived from a Circular Peer-Review consensus pipeline, then splits those rubrics into Must-Right checks for essential facts and Easy-Wrong checks for fine-grained details. The dual-stream design exposes brittleness that holistic semantic matching scores consistently hide. Teams selecting vision models for production should run Easy-Wrong audits before trusting aggregate benchmark numbers. link

06 [Theory] The State-Prediction Separation Hypothesis The standard transformer conflates two distinct jobs in one computation stream: predicting the next token and storing state for future predictions. Separating those roles into two dedicated streams yields lower validation loss at matched parameter counts across multiple pretraining scales, suggesting the single-stream design allocates capacity inefficiently between the two functions. The gains are consistent, not scale-dependent, which makes this a structural finding rather than a tuning result. Practitioners designing new architecture variants have a testable hypothesis: dedicated state capacity is worth the added complexity. link