Also Worth Noting - 2026-07-31

Also Worth Noting

02 [Agent] Metis: Memory Foundation Model Agent memory has always been bolted on as an external retrieval module. Metis internalizes memory as a native model capability, treating it as a persistent, dynamically evolving state rather than a plugin the model queries at inference time. The architectural shift means an agent can update and carry knowledge across interactions without a separate vector store in the loop. Teams building long-horizon agents who want to reduce external infrastructure dependencies should watch this direction closely. link

03 [Inference] Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes Relaxing distributional matching in speculative decoding to gain speed silently rewrites the output distribution in ways that standard benchmarks rarely surface. Lossy verification schemes accept draft tokens that strict rejection sampling would discard, shifting token probabilities enough to change downstream task behavior in measurable, sometimes severe ways. The speed gains are real; so is the accuracy tax. Teams running speculative decoding in production should audit generation quality on task-specific distributions, not just throughput numbers. link

04 [Training] Beta-OPSD: Deriving with Policy Optimization, Training with Self-Distillation Vanilla on-policy self-distillation for reasoning models is brittle because it implicitly fixes the KL penalty weight at exactly 1, leaving no room to trade off stability against exploration. Reframing OPSD as the beta=1 member of a broader policy-optimization family turns that implicit constant into a controllable regularization parameter. Sweeping beta recovers training stability without additional engineering scaffolding. For teams training reasoning models with self-distillation, this one-parameter generalization may be the cheapest path from unreliable to reliably trainable. link

05 [Eval] Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions Deep research agents accept fabricated but credible-looking sources at high rates when those sources are seeded into their retrieval context. MisKnow-Agent constructs task-specific false documents with controlled authority cues and source styles, generating 5,933 filtered misleading documents drawn from DeepResearch Bench tasks. The evaluation framework reveals a live reliability risk: any deployment where adversaries can influence the retrieval corpus is exposed today, not in some future threat model. Teams shipping deep research pipelines over open or user-contributed corpora should treat this as a current production concern. link

06 [Theory] Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations Demographic bias in LLMs concentrates in a locatable neuron subset inside GLU-MLP layers, not distributed uniformly across the network. Fairness Pruning identifies those neurons by diffing activations on minimally contrastive prompt pairs and reading the signal at the down_proj input, tested on models up to 3 billion parameters. Surgical removal of the flagged neurons offers a cheaper path to bias mitigation than full retraining or fine-tuning. For teams managing bias in production models where retraining is expensive, this localization step alone may be worth running as a diagnostic. link