Also Worth Noting - 2026-08-02

Also Worth Noting

02 [Training] Reusing Rollouts under Policy Lag: Prefix-Normalized Policy Optimization for LLM Reinforcement Learning Standard PPO discards each rollout batch after a single learner update, burning generation compute on every step. Prefix-Normalized Policy Optimization corrects for off-policy drift by normalizing importance weights at the prefix level rather than accumulating a product across all token positions, which otherwise produces an unmanageable dynamic range. This lets multiple learner updates run safely against the same rollout batch. Teams running LLM RL at scale can cut generation cost without the divergence that kills performance on later updates. link

03 [Agent] V-Mem: Modality-Routed Retrieval for Long-Term Multimodal Agentic Memory Multimodal agent memories fail on vision queries not because retrieval is generally weak, but because text and image embeddings occupy incompatible index spaces, so similarity search silently returns wrong-modality results. V-Mem routes queries by modality before search, ensuring a vision question never competes against text embeddings in the same index. The fix requires no retraining of the underlying model. Any team storing mixed-modality conversation history in a vector index is exposed to the same silent failure this paper diagnoses. link

04 [Inference] Attend to Your Own Thoughts: Breaking the Barrier for Post-Training Quantization of Reasoning LLMs through the Lens of 1.58-Bit Quantization Ternary post-training quantization collapses on math and coding benchmarks not because 1.58-bit is too aggressive, but because standard calibration schemes ignore chain-of-thought tokens entirely. ScaleQ-1.58 treats intermediate reasoning tokens as first-class attention targets during calibration, which is the fragile point prior methods miss. Performance recovers at 1.58-bit without any retraining. Teams quantizing reasoning models for edge or cost-constrained serving should audit whether their calibration data includes CoT traces. link

05 [Theory] No Single Neuron of Failure: Distributed Safety Alignment Against White-Box Attacks Concentrating safety behavior in a small neuron cluster creates an exploitable single point of failure: white-box attackers targeting open-weight models can disable alignment by ablating just that cluster. Distributed Safety Alignment redundantly encodes safety signal across many neurons, raising the cost of any neuron-level attack substantially. Measured results on ablation attacks confirm the safety baseline holds even when critical individual neurons are removed. For teams releasing or deploying open-weight models, sparse safety representations are an architectural liability worth addressing at training time. link

06 [Eval] Loud or Silent? A Reusable Framework for Per-Modality Failure Analysis in Multimodal Clinical AI Clinical multimodal models evaluated with all modalities present routinely fail silently when one goes missing at deployment, and standard accuracy metrics cannot distinguish which modality caused the drop or whether the failure is even detectable. This model-agnostic framework answers both questions at the per-example level, separating loud failures the system can flag from silent ones it cannot. The evaluation is designed to be reused as models are swapped out. The distinction between loud and silent failure applies directly to any production multimodal system, not only medical ones. link