Also Worth Noting - 2026-07-04

Also Worth Noting

02 [Agent] AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents Appending all reflections to every prompt hurts more than it helps across most long-horizon tasks. AgenticSTS replaces the standard cross-decision transcript with a bounded contract: each decision sees only a fresh message assembled by typed retrieval, keeping prompt size flat regardless of run length. That isolation lets individual memory components be toggled and measured causally. Reflection entries, treated as a universal good in current practice, turn out to be net negatives in the majority of tested tasks. link

03 [Eval] PACE: A Proxy for Agentic Capability Evaluation A single SWE-Bench run can cost thousands of dollars and take days to complete. PACE shows that a small, carefully selected subset of non-agentic atomic capability scores predicts agentic benchmark rank with enough fidelity to make full runs optional for most development decisions. The proxy composite covers reasoning and code generation tasks that are fast and cheap to run. Teams iterating on agent architectures can screen candidates with PACE before committing to expensive infrastructure-heavy evals. link

04 [Training] Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning Outcome-only RL on clinical image tasks lets early reasoning errors compound silently through the chain, producing failure cascades that final-answer supervision cannot detect or correct. Step-aware credit assignment targets each intermediate reasoning step directly, giving the optimizer a signal before errors propagate. The result is a measurable reduction in cascade rate and improved intermediate reasoning correctness on medical VQA benchmarks. For teams building clinical AI pipelines, this argues strongly against treating the final label as the only training signal. link

05 [Inference] Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling MrFlow matches full-resolution generation quality at one-fifth the compute, with no retraining and no custom kernels required. The method stages upsampling across the diffusion trajectory, generating at low resolution early in the process and stepping up only as the trajectory converges, which avoids the blurring artifacts that plagued earlier multi-resolution approaches. The 5x speedup claim holds on existing model checkpoints without any distillation step. Drop-in deployment on current text-to-image pipelines is the immediate practical payoff. link

06 [Theory] Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads Current retrieval-head detectors are systematically blind to any attention head that paraphrases rather than copies, because they reward heads whose attended token matches the generated token literally. Logit-contribution scoring shifts the criterion to what a head writes through its output-value circuit, catching heads that synthesize meaning without reproducing surface tokens. The result is that existing long-context attribution maps flag the wrong layers. Teams using mechanistic interpretability to audit long-context model behavior should treat prior retrieval-head inventories as incomplete. link