Also Worth Noting - 2026-06-26

Also Worth Noting

02 [Inference] JetFlow: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting Speculative decoding hits a throughput ceiling not because draft budgets run out, but because longer draft paths lower acceptance rates and causality constraints force sequential overhead. JetFlow breaks that link by drafting across tree branches in parallel, decoupling draft budget from causal depth. Where head-based methods plateau, parallel tree drafting keeps yielding throughput gains. Teams running high-draft-budget speculative decoding and hitting diminishing returns should look at this architecture before assuming the ceiling is fundamental. link

03 [Theory] When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models No multi-model ensemble strategy can exceed one minus the rate at which every member model fails on the same query. That single quantity, the co-failure rate, caps routing, voting, cascades, and mixture-of-agents alike, yet almost no evaluation reports it. Average pairwise error correlation, the metric teams do report, cannot detect this ceiling because two error distributions with identical marginals and pairwise correlations can have very different all-wrong rates. Before adding a fifth model to an ensemble, measure co-failure rate first. link

04 [Agent] Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It RL-trained tool-use agents sometimes abruptly stop invoking tools correctly mid-training, not because the underlying capability disappears, but because probability spikes on specific control tokens disrupt structured execution. The capability is still there; the formatting scaffold collapses around it. Adding step-level supervisory signals stabilizes those control token distributions and recovers the gains RL was supposed to deliver. Teams hitting sudden tool-invocation failures during RL fine-tuning now have a concrete mechanism to diagnose and a concrete fix to apply. link

05 [Eval] OpenBioRQ: Unsolved Biomedical Research Questions for Agents Agentic models almost never produce broken citation links, over 99% resolve, but roughly 15.9% of those citations point to a real paper that does not support the claim being made. Fixed-answer-key benchmarks cannot catch this because models can reproduce the expected source directly from the answer key without verifying the source actually backs the claim. OpenBioRQ uses 12,553 unsolved biomedical questions to close that gap. Any team evaluating agentic retrieval on biomedical tasks should treat citation validity, not just citation resolution, as a first-class metric. link

06 [Training] Information-Aware KV Cache Compression for Long Reasoning Attention-weight-only KV cache compression drops tokens that are contextually redundant but carry high predictive uncertainty, exactly the tokens long reasoning chains depend on. Adding entropy-based signals, what the paper calls Forward Influence, measures how much each token shapes future predictions rather than how much past context attends to it. The result is smaller cache footprints on long reasoning traces without the accuracy degradation that attention-only pruning incurs. Teams compressing KV caches for extended chain-of-thought workloads should audit whether their importance metric is backward-looking or forward-looking. link