Also Worth Noting - 2026-08-22

Also Worth Noting

02 [Eval] MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use Correctly retrieved, semantically relevant memories can still degrade model reasoning on the current task. Standard memory benchmarks check whether retrieval succeeded; MemTrapBench checks what happens after, exposing a class of memory-induced cognitive traps that correct retrieval scores mask entirely. The benchmark targets failure modes where the model's beliefs or reasoning chains are distorted by what it faithfully remembered. Teams shipping long-term memory into production should audit against this failure class, not just retrieval accuracy. link

03 [RAG] Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization Parametric knowledge internalization can match retrieval-augmented QA on bounded corpora, removing the inference-time retrieval hop entirely. IAR separates the process into three stages: a structured injection step that converts documents into continuation-style training signal, a QA alignment step, and a recovery step that restores general capabilities eroded by domain-specific fine-tuning. The staged separation is what makes it work; collapsing the stages into conventional continued pretraining does not. For latency-sensitive deployments where a retrieval round-trip is too costly, this is a direct alternative path. link

04 [Inference] FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving Sparse prefill attention has been theoretically attractive for years; FlashPrefill V2 is the first version close enough to production to matter. The original FlashPrefill used instantaneous pattern discovery and max-based dynamic thresholding to skip attention blocks, but stayed at prototype distance from real serving infrastructure. V2 closes that gap across three engineering dimensions, delivering measurable prefill speedups for the long-context regimes where quadratic attention is today's actual bottleneck. Infrastructure teams running 32K-plus context workloads should track this implementation. link

05 [Application] Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis Single-answer evaluation systematically penalizes correct retrosynthesis predictions that happen to be alternative valid reaction paths. Top-K prompting reframes training and inference to capture diverse plausible predictions rather than forcing a single-answer race, and the C3LM model trains on CREED-CCV-2+USPTO-XL, a dataset of roughly 45.6 million verified reactions. ChemCensor-based and novelty-oriented rewards push the model toward chemically plausible outputs rather than benchmark-optimized ones. The fix is a training paradigm shift, not a scale increase. link

06 [Agent] PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents Blocking forbidden actions is not sufficient for policy compliance: agents also fail by skipping required procedural steps, a failure mode that action-local runtime safeguards cannot detect. PolicyGuide compiles each domain policy into a workflow representation and guides the agent through the full procedure, catching omission failures alongside commission ones. Existing workflow-following systems target completion, not compliance; existing guardrails target individual actions, not procedural coverage. Customer-service deployments that rely on action-level checks alone are exposed to the omission failure class this paper formalizes. link