Also Worth Noting - 2026-08-21

Also Worth Noting

02 [RAG] Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization Fine-tuning a model on a document corpus quietly breaks its general abilities, a tradeoff most retrieval-free approaches ignore. IAR separates this into three explicit stages: structured injection converts source documents into continuation-style training, alignment shapes QA behavior, and a recovery stage repairs the general-capability degradation that injection causes. The staged separation is the design choice, not a training trick. Teams considering retrieval-free deployments where latency or privacy rules out a vector store should treat the recovery stage as a required component, not an optional cleanup pass. link

03 [Inference] FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving Prefill is 3-5x more expensive than decode for long-context requests, and most sparsity approaches stay at the prototype stage. FlashPrefill V2 moves block-sparse prefill attention toward production along three dimensions, combining instantaneous pattern discovery with max-based dynamic thresholding at serving scale. The gap between algorithmic prototype and deployable system is where most long-context efficiency work quietly stalls. Infrastructure teams running long-context serving should watch this as a concrete path to cutting prefill costs without changing model weights. link

04 [Agent] PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents Action-local safety filters structurally cannot catch procedural omissions, such as a customer-service agent skipping identity verification before processing a change. PolicyGuide compiles each domain policy into a workflow-level representation, shifting compliance checking from single-action guards to full procedure orchestration. The distinction matters because skipped steps and forbidden actions are different failure classes, and only one of them is visible to a per-action filter. Teams deploying customer-service agents under regulatory or organizational policy constraints should audit whether their current guardrails cover omission failures at all. link

05 [Eval] Towards Quantifying Benchmark Optimization in ASR Models ASR leaderboard scores can inflate without any improvement in real-world transcription, because models can learn to reproduce benchmark reference spans even when the audio does not clearly determine the correct output. Three families of behavioral probes expose this: reference disagreement, masked-number recovery, and a third class targeting cases where audio underdetermines the transcript. The methodology quantifies benchmark overfitting rather than just asserting it exists. Teams selecting ASR models for production should run at least the reference-disagreement probe before trusting public benchmark rankings. link

06 [Application] TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity 146,505 parameters with no attention mechanism produces competitive probabilistic zero-shot forecasting, by offloading period detection entirely to a zero-parameter spectral step rather than learning it from data. A spectral detector supplies dominant periods, the context is folded on their phase, and a dilated convolutional encoder with a block-autoregressive quantile decoder handles the rest. TinyCast sits at the size-accuracy frontier on GIFT-Eval among zero-shot entries whose parameter counts are publicly verifiable. For edge deployments or cost-sensitive forecasting pipelines, this challenges the assumption that competitive zero-shot time-series modeling requires a large pretrained backbone. link