Also Worth Noting - 2026-08-16

Also Worth Noting

02 [Inference] KV-Rescue: Recovering Reasoning Language Model KV Eviction Loss via Stepwise Interleaving Aggressive KV-cache eviction on long reasoning traces causes not just accuracy drops but runaway degeneration, where the model produces incoherent or repetitive tokens until it hits the length limit. The root cause is an information gap, not a capability gap: an evicted 7B model and a full-context 1.5B model make complementary errors, meaning the context is missing, not the skill. Stepwise interleaving patches that gap without expanding the cache budget. Teams running long chain-of-thought inference under tight memory budgets should treat degeneration as a retrieval problem first. link

03 [Hardware] Large Models for Small Devices: Recent Advances and Empirical Analysis of Edge AI Deployment What compresses well on paper frequently does not deploy well on real silicon. This survey of dozens of compression studies extracts deployment guidelines from actual GPU, CPU, and Raspberry Pi runs, and finds that compression ratio and real-hardware throughput diverge by a wide margin across tasks. No single technique wins across all settings: for question answering, Qwen3.5 0.8B reaches 93.85 SQuAD F1 under Q5_K_M quantization, but image segmentation tells a different story. Anyone benchmarking compression methods without measuring on target hardware is optimizing the wrong metric. link

04 [RAG] Navigation-Informed Embeddings: Dense-Retriever Adaptation from Agent Search Traces Agent retrieval workflows already produce query, retrieval, and stopping traces as a byproduct of answering questions, and those traces alone are enough to adapt a dense retriever to a shifting workflow distribution. Navigation-Informed Embeddings (NIE) turns the stopping document into a soft positive and uses preceding path documents as hard comparisons with ordinal constraints and geometric decay, requiring no new relevance labels, synthetic queries, or LLM judgments. Teams maintaining agentic RAG pipelines can cut a recurring labeling cost to near zero by treating their own search traces as a continuous fine-tuning signal. link

05 [Agent] Bounded Agents: Delegation Security for Multi-Agent AI Systems Static permissions set at session start let agents combine individually permitted actions into outcomes the user never authorized, a compound-action problem that prompt injection mitigations do not address. The authorization gap runs deeper: agents can also delegate authority to sub-agents without narrowing it, expanding the blast radius with each hop. Bounded delegation formalizes a fix by constraining what authority can be passed downstream and scoping each request against prior actions in the session. Anyone deploying multi-agent tool-use today should audit whether their permission model handles compound outcomes, not just individual action approvals. link

06 [Application] Learning Auditable Classifier Models: Source-Disjoint Tree Ensembles Standard boosted ensembles couple structure discovery with coefficient estimation in a way that makes compact per-prediction auditing intractable, a real problem in clinical and regulated settings where inspecting individual decisions is a compliance requirement. Residual Pattern Tree Ensemble (RPTE) breaks that coupling through a three-stage source-disjoint splitting process, producing audit trails that regulators can inspect without the accuracy penalty typically imposed by generalized additive models. GAMs restrict interactions to pairwise terms; RPTE does not, and it still delivers audit trails at the per-prediction level. Teams building models for high-stakes tabular applications should track this as a practical alternative to post-hoc rule extractors. link