Also Worth Noting - 2026-06-30

Also Worth Noting

02 [Inference] ReFreeKV: Towards Threshold-Free KV Cache Compression Most KV cache pruning methods carry a hidden cost: someone has to pre-tune a budget threshold per input domain, and picking the wrong one silently degrades quality. ReFreeKV removes that requirement entirely by adapting the compression budget at inference time without domain-specific calibration. The result is lossless compression across open-domain inputs that vary in length, topic, and difficulty. Teams running LLM serving across mixed-domain traffic can drop the per-deployment threshold search that makes most KV pruning impractical outside controlled benchmarks. link

03 [Agent] Agentic Abstention: Do Agents Know When to Stop Instead of Act? Current LLM agents fail at a specific, underexplored failure mode: continuing to call tools on goals that are unachievable or underspecified, rather than stopping. Unlike single-turn abstention, agentic abstention is a sequential decision problem across search, browsing, and terminal tools, where each additional step compounds the failure. This paper is the first to formally define and benchmark that abstention gap. For teams deploying agents in production, the practical implication is direct: tool-call budgets and stopping criteria deserve as much design attention as tool selection. link

04 [Eval] Beyond IID: How General Are Tabular Foundation Models, Really? Tabular foundation models score well on standard benchmarks because those benchmarks are mostly drawn from distributions where these models already excel. Cross-discipline evaluation, pulling tasks from domains outside the training distribution, exposes systematic generalization failures that discipline-specific leaderboards structurally cannot surface. Benchmark software and evaluation protocols remain fragmented across research communities, which means model developers rarely see the failure signal. Teams adopting tabular foundation models for domain-specific applications should run out-of-distribution evaluation before trusting in-distribution leaderboard rankings. link

05 [Training] AsyncOPD: How Stale Can On-Policy Distillation Be? On-policy distillation can tolerate one to two steps of gradient staleness before student quality measurably degrades. That tolerance window is wide enough to fully decouple rollout generation from learner updates in an async pipeline, eliminating the on-policy systems bottleneck that makes reasoning workloads expensive. Prior work studied stale data in async RL but left the OPD case open; this paper closes it with direct measurements. Teams running LLM post-training for reasoning tasks can adopt async pipelines without paying a quality penalty, as long as staleness stays within that window. link

06 [Theory] How Good Can Linear Models Be for Time-Series Forecasting? Tuning normalization, differencing, and context length in Ridge regression closes most of the accuracy gap to large time-series transformers, without scaling model capacity at all. Ridge regression has a closed-form solution and interpretable weights, which means the optimal preprocessing configuration can be read directly from the hyperparameter search rather than inferred from black-box ablations. The finding suggests the field has been scaling models to compensate for poor preprocessing choices rather than genuine capacity limits. Teams evaluating time-series foundation models should run a well-tuned linear baseline before committing to transformer-scale infrastructure. link