Also Worth Noting - 2026-07-26

Also Worth Noting

02 [Training] From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement RLVR has been boxed into math and code because verification requires ground truth. RLSVR breaks that ceiling by reformulating open-ended tasks as self-verifiable problems, borrowing the pretext-task logic from self-supervised learning to manufacture deterministic reward signals without a reward model or LLM judge. The approach removes evaluation bias and the inference cost of an external judge in one move. Teams looking to apply verifiable-reward RL to writing, planning, or summarization now have a concrete construction path. link

03 [Agent] ACM: Agentic Context Management for Long Horizon Tasks Fixed token-count triggers for context compression fire at the wrong moment , they ignore what the agent is actually reasoning about when they flush state. ACM gives agents purpose-built context editing tools and lets them decide autonomously when and what to compress, modeled on the interplay between short-term and long-term human memory. The result is lossless compression aligned with the agent's evolving reasoning focus rather than a heuristic threshold. Any multi-step agent regularly exceeding 32k tokens should treat this as a direct replacement for count-based compression. link

04 [Inference] Kalypso: Relational LLM Serving Every row in a filtered table hitting an LLM as an independent request wastes the structure the query planner already knows. Kalypso exposes relational query semantics to the serving layer, enabling pipelined execution across semantic operators so intermediate tuples flow directly into downstream batches before a full result set materializes. Batching and caching opportunities that are invisible to request-centric systems become actionable. If your data pipeline calls an LLM for filtering, extraction, or ranking at table scale, Kalypso shows the current per-row serving model is leaving significant latency on the table. link

05 [RAG] TriShieldRAG: A Three-Ring Defense-in-Depth Framework Against Knowledge Corruption in Retrieval-Augmented Generation PoisonedRAG showed that as few as five crafted documents flip an undefended system to a chosen wrong answer , yet most RAG hardening work targets query-time prompt injection, not the write path. TriShieldRAG addresses the specific threat of adversarial document insertion into writable knowledge bases with a three-layer defense covering ingestion, retrieval, and generation. It is the first systematic treatment of this vector rather than a point patch. Teams running RAG over knowledge bases with multi-party write access should audit their ingestion pipeline against this framework before the retrieval layer. link

06 [Training] Hybrid Advantage Estimation with Unified Critic for VLM Agentic Reinforcement Learning Token-wise RL over concatenated multi-turn trajectories dilutes turn-level credit; turn-wise RL with uniform within-turn credit loses token-granularity signal. HAUC derives a hybrid advantage estimator that satisfies both objectives simultaneously, backed by theoretical formulations for each optimization level. A single unified critic carries both signals without adding a second model. The approach delivers measurable gains on multi-turn agentic benchmarks and is directly useful for teams training VLM agents on interactive tasks where credit assignment across turns has been the persistent weak point. link