Also Worth Noting - 2026-07-28

Also Worth Noting

02 [Inference] Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification Fixed-budget sparse attention misses the point for video diffusion transformers: token importance varies too much across frames for any rigid block quota to stay accurate. Sol-Attn selects key-value blocks on the fly by targeting a cumulative proxy probability mass rather than a fixed top-k count, letting the budget expand or contract per head. The routing itself runs without any model retraining or architecture change. Teams running DiT-based video pipelines can drop this in as a training-free inference optimization. link

03 [Agent] From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search Outcome-based RL for agentic search gives only sparse supervision, one signal per completed task, leaving the model to guess which reasoning steps actually mattered. Multi-agent protocol distillation sidesteps the logit-matching problem entirely by having proprietary teacher models produce structured reasoning protocols that open-source students can imitate step by step. The approach yields measurably denser supervisory signal than RL alone across multi-hop retrieval tasks. For teams trying to close the gap between open and closed models on knowledge-intensive agents, this is a more tractable path than scaling RL reward density from scratch. link

04 [Application] StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents Two completely different application states can render as identical screenshots, which means perception-first computer-use agents are working from an ambiguous signal by design. StateAct routes around this by making a code-writing main agent operate directly on files, DOM, and application backends, while a dedicated GUI subagent handles only the interactions that genuinely require pixel-level vision. The split is not cosmetic: bypassing the lossy screenshot representation cuts errors on long-horizon tasks where state divergence is invisible to the eye. Teams building computer-use pipelines should treat program state access as a first-class capability, not a fallback. link

05 [Hardware] Characterizing Warp Divergence from Pascal to Blackwell Warp divergence handling changed silently across GPU generations after Volta, meaning assumptions baked into kernels on Ampere do not transfer cleanly to Hopper or Blackwell. Across all tested architectures, divergent paths serialize linearly with path count k, following T(k) approximately sk with no super-linear reconvergence penalty, but the constant s shifts between generations in ways that cycle-accurate microbenchmarks expose. Static analysis of compiler-generated SASS reveals additional architectural differences invisible to high-level profiling. Kernel authors tuning divergence-sensitive code on Ampere should re-benchmark on Blackwell before assuming the optimization holds. link

06 [Eval] Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models Longer reasoning traces actively degrade hallucination detectors, not because the model reasons poorly, but because irrelevant and repetitive steps bury the truthfulness signals that scoring methods depend on. Reasoning Denoiser identifies two noise types, irrelevant steps and repetitive steps, and filters them before any confidence or embedding-based scoring runs. Removing the noise raises detection accuracy, which means trace length is a poor proxy for reasoning quality. Anyone deploying chain-of-thought systems and relying on trace-based confidence scores should treat trace cleaning as a preprocessing step, not an afterthought. link