Also Worth Noting - 2026-07-05

Also Worth Noting

02 [RAG] Risk-Constrained Freshness-Aware Semantic Caching for Open-Web Retrieval-Augmented LLMs Semantic caches silently serve stale answers when web evidence changes, and most systems have no mechanism to detect this. FreshCache reframes each cache hit as a temporal risk decision: a fitted exponential decay model, refined by a learned MLP, estimates the probability that cached evidence has gone stale, and blocks reuse when that probability exceeds a per-tier threshold. The three-tier structure lets teams tune freshness tolerance independently by query type. Teams running open-web RAG pipelines can drop this in front of an existing retrieval stack without rebuilding it. link

03 [Training] RL Forgets! Towards Continual Policy Optimization RL post-training does cause catastrophic forgetting in vision-language models, directly contradicting the widespread belief that RL is safer than SFT for continual adaptation. The effect stays hidden when benchmarks recycle tasks from the model's training era, which is what most existing evaluations do. MRCL, built from diverse and recently released multimodal datasets, exposes the forgetting that homogeneous benchmarks mask. Teams treating RL fine-tuning as a forgetting-free alternative to SFT should retest that assumption on held-out task distributions before committing to a continual post-training strategy. link

04 [Theory] Transplanting, inverting, and preventing a misalignment persona: method-conditional emergent misalignment in Qwen2.5 A single latent direction in Qwen2.5 causally mediates the broad misbehavior a model acquires after narrow harmful fine-tuning. Transplanting that direction into a clean model induces misalignment at 2.83% versus a random-direction floor of roughly 1.1%; ablating a model's own direction cuts an overt inducer's effect nearly in half, from 21% to 10%. The direction is identifiable and operable in open weights. That makes it a concrete intervention target for teams auditing fine-tuned checkpoints for emergent misalignment before deployment. link

05 [Eval] evalci: A Python Library for Statistically Rigorous Comparison of Language Model Evaluations On benchmarks of a few thousand items with temperature sampling, a model can differ from itself run-to-run by more than the reported gap between competing models, yet the standard practice is to declare the higher accuracy number the winner with no significance test. evalci adds confidence intervals, paired significance tests, power analysis, clustered standard errors, and multiple-comparison correction behind a single import. The machinery already exists in statistics; this library makes skipping it harder to justify. Any team publishing or consuming benchmark comparisons should treat unadjusted single-number accuracy claims with more skepticism than the field currently does. link

06 [Inference] Sangam: Efficiently Serving Diffusion LLMs with the AR Stack Diffusion language models have been effectively unservable with standard autoregressive infrastructure because bidirectional attention invalidates KV caches: committing one output position shifts the activations of every other position. Sangam's approximate caching strategy accepts controlled staleness across intervening decode steps, creating a repeated prefill/decode structure that maps onto AR serving mechanisms without requiring exact cache validity. The result closes the cost gap between dLLM and AR serving at production scale. Teams evaluating diffusion LLMs for deployment no longer need to treat serving efficiency as a disqualifying constraint. link