Also Worth Noting - 2026-08-09

Also Worth Noting

02 [Agent] BOUND: Brief-Guided Corrective Preference Distillation at Search-Control Boundaries Outcome-level supervision for deep search agents cannot tell the difference between a task-aligned continuation and a locally plausible one that locks in the wrong anchor. BOUND fixes this at decision-time boundaries: for each student-induced state, a teacher-side search-state brief captures global task constraints, and preference distillation is applied only where drift is most likely to take hold. The result targets wrong-anchor drift, constraint drift, and local-topic drift as distinct failure modes rather than collapsing them into a single trajectory loss. Teams building multi-hop retrieval agents should treat boundary-level preference signals as a separate supervision target from outcome reward. link

03 [Inference] DistillCache: KL-Guided Adaptive KV-Cache Eviction for Memory-Efficient LLM Inference H2O and SnapKV drop KV-cache tokens by attention score or position, missing tokens whose future predictive influence is high but whose past attention weight is low. DistillCache reframes eviction as a sequential decision problem and trains a lightweight policy network using KL divergence between full-cache and evicted-cache output distributions as the reward signal. That KL gap is a direct measure of how much a dropped token changes what the model predicts next, not a proxy. At equivalent memory budgets, DistillCache recovers measurable perplexity that heuristic eviction methods leave on the table. link

04 [Training] Deployable Per-Instance Multi-Layer Activation Steering for Large Language Models Fixing injection layers globally per task is the standard activation steering default, but on two open-weight 8B models across six binary persona traits, a per-instance oracle over layer subsets shows the best layers shift from input to input. No single global layer set recovers the per-instance oracle benefit on most trait-model pairs. A greedy selection rule closes much of that gap without weight updates to the frozen model. Teams using activation steering for persona or behavior control should treat layer selection as an instance-level decision, not a one-time tuning choice. link

05 [Eval] Can Open-Weight Models Compete on Financial Text Comprehension? Chinese open-weight models closed the benchmark gap against proprietary frontier models in recent months, but those benchmarks rarely test document-grounded financial reasoning. The updated Financial Touchstone benchmark covers 2,967 question-context-answer triplets drawn from 495 real international annual reports, expanding model coverage from eleven to twenty models across ten providers, including GLM 4.7, GLM 5, Kimi K2.6, and DeepSeek V3.2. Whether the benchmark gap closure holds when the task is extracting precise figures from dense annual report prose is exactly what this update is designed to answer. Teams deploying open-weight models for financial document workflows have a more grounded comparison surface now. link

06 [Theory] Learning from Consensus and Disagreement: Unsupervised On-Policy Self-Distillation with Minority-Trajectory Contrast Every competitive on-policy distillation method today requires gold solutions or an external verifier to give the teacher an information advantage over the student. CoDA drops both by treating rollout consensus as a proxy for correctness: trajectories that the model's own unlabeled rollouts agree on form the reliable signal, and minority trajectories that diverge from consensus are contrasted against them to construct teacher privilege without ground truth. The framework is fully unsupervised and operates entirely on the model's own latent uncertainty structure. This matters most for domains where annotated solutions are scarce or where no reliable verifier exists. link