Also Worth Noting - 2026-08-13

Also Worth Noting

02 [Agent] SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries Skill libraries are growing faster than context windows can accommodate, and the standard fix of retrieving whole skill packages is the wrong unit of reuse. SkillZip compresses skill graphs at the sub-skill level while preserving procedural contracts, so the executable structure and downstream tool call signatures survive compression intact. The result is more capability packed into a fixed context budget without breaking the routines that depend on those contracts. Teams managing large agent skill registries should treat this as a concrete alternative to naive retrieval-and-truncate approaches. link

03 [Training] Parameter Exploration for RLVR via Variational Learning Temperature scaling can widen the output distribution but cannot reorder token probabilities, which means exploration ceilings in RLVR are partly an architectural artifact of where exploration happens. This method moves exploration into parameter space, sampling rollouts from weight perturbations rather than output logits, which allows the model to surface token orderings that temperature-based methods structurally cannot reach. For practitioners hitting stalled training or divergence in GRPO-style recipes, parameter-space exploration is a concrete lever that action-space methods leave untouched. link

04 [Inference] ReRound: Reconstructive Rounding to Resolve Midpoint Ambiguity in Calibration-Free LLM Quantization Round-to-nearest quantization has a known blind spot: weights sitting near the center of a quantization interval carry no signal about which direction to round, and that ambiguity quietly degrades accuracy. ReRound trains a conditional diffusion model to produce continuous reconstructions of low-bit weights, then uses those reconstructions as a guidance signal to break the tie at midpoints. No calibration dataset is required. Teams deploying calibration-free INT4 quantization gain a targeted fix for one of the cleaner accuracy gaps in the standard RTN pipeline. link

05 [Eval] The Embedder's Dilemma: LLMs Are Better, but at What Cost? Across 37 tasks the best LLM embedder (Gemini 3.1 Pro, 77.6) and the best dedicated embedding model (77.2) differ by 0.4 points , a statistical tie. The task-level breakdown matters more than the aggregate: LLMs lead on reasoning-heavy retrieval, while dedicated models lead on classification. That 0.4-point parity comes at 10-100x higher per-query cost for the LLM path. For most production retrieval workloads, the economics make switching unjustifiable unless the workload skews heavily toward multi-step reasoning queries. link

06 [RAG] RAGSieve: Self-Referenced Local Contrast for Knowledge-Poison Detection in Retrieval-Augmented Generation Most RAG poison detectors require either a trusted clean reference corpus or attack-specific signatures, both of which are unavailable in production deployments. RAGSieve builds its reference from the inspected system itself: the query-local variant scores the top-five retrieved candidates against ranks 6-20 of the same retrieval, detecting answer-anchor concentration and carrier transitions without any external ground truth. A graph-based complement catches structural anomalies across the retrieved set. Teams running open-corpus RAG pipelines get a detection path that does not depend on a clean baseline they cannot guarantee. link