Also Worth Noting - 2026-07-08

Also Worth Noting

02 [Inference] Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling Prior chunk-wise sparse attention methods fall short of full attention not because sparsity is wrong, but because their chunk selection is inaccurate. HiLS Attention fixes this by introducing hierarchical landmark tokens that let each query identify which chunks actually matter before computing attention, learning the selection end-to-end under the language-modeling loss rather than relying on heuristics. The result is sparse compute that matches full attention quality. Teams running long-context inference at scale now have a chunk-wise method worth taking seriously as a drop-in replacement. link

03 [Training] TREK: Distill to Explore, Reinforce to Refine GRPO stalls on hard prompts not because the optimizer is weak, but because correct solution modes sit entirely outside the student's on-policy sampling distribution. TREK repurposes distillation as an exploration tool: forward-KL distillation seeds the student's distribution with solution trajectories from a teacher, expanding coverage to modes the student would never find on its own, then hands off to GRPO once those modes are reachable. Because TREK only consumes verified output trajectories, the teacher can be a black box, a white box, or the same model given extra inference compute. Teams hitting hard-prompt failure walls with GRPO should treat this staged approach as a concrete fix. link

04 [Inference] Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding A single checkpoint can now serve as three different decoders depending on workload. Nemotron-Labs-Diffusion trains with a joint AR-diffusion objective, where diffusion improves lookahead planning and AR supplies left-to-right linguistic priors, and the two objectives turn out to be complementary rather than competing. In self-speculation mode, diffusion drafts while AR verifies, outperforming multi-token prediction methods. Operators get AR mode for low-concurrency latency, diffusion mode for high-concurrency throughput, and self-speculation as a third gear, all without retraining or swapping checkpoints. link

05 [RAG] When Classic Cache Policies Fail: Learning-Augmented Replacement for Semantic Retrieval Buffers LRU and LFU consistently underperform naive FIFO on semantic retrieval workloads, a result that holds across eight replacement policies tested on two MemoryBench-Full datasets. The failure is structural: classic eviction policies assume binary cache hits, but semantic buffers match by embedding similarity, making hit quality continuous rather than binary. A learned replacement policy exploits that continuous structure and cuts retrieval latency where LRU and LFU bleed it. Any team using retrieval buffers in production agents should audit their eviction policy before assuming the cache is helping. link

06 [Agent] SkillOpt-Lite: Better and Faster Agent Self-evolution via One Line of Vibe The elaborate multi-stage pipelines in prior agent skill-optimization work are not necessary. SkillOpt-Lite maps skill optimization onto zeroth-order gradient estimation, drawing direct parallels to classical central-difference and trust-region methods, and shows that a single update step matches or beats complex pipelines on standard benchmarks. Skill trajectories act as interpretable debugging feedback, giving the optimizer signal that blind numerical perturbations in classical zeroth-order methods never had. For teams building self-evolving agents, this is a strong argument for stripping the pipeline back before adding components. link