Also Worth Noting - 2026-08-11

Also Worth Noting

02 [Inference] OasisKV: Scaling In-Decode KV Cache Beyond HBM with Lookahead Sparse Prefetching HBM scarcity, not compute, is now the binding constraint for long-context LLM serving. OasisKV decouples full KV cache storage from HBM by predicting which sparse keys will be needed before each decode step, then prefetching them from DRAM or SSD just in time. The lookahead sparsity signal lets the system hide most of the off-chip latency, recovering near-HBM throughput at 4x the context length. Teams running 128k+ token workloads on memory-constrained hardware have a concrete offloading path here. link

03 [Training] On-Policy Self-Distillation without Any Supervision On-policy self-distillation with zero external labels, no ground truth, no teacher model, no environment feedback, turns out to be achievable using only internal consistency across a model's own rollouts. U-OPSD samples multiple generations, constructs a pseudo-solution by comparing them for agreement, and uses that signal to update the policy on-policy. The method cuts the dependency on annotated data or a larger teacher that nearly every post-training pipeline currently assumes. Worth watching for settings where labeling is expensive or a capable teacher is unavailable. link

04 [Agent] Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory Small LLMs fail at multi-step agent tasks partly because they cannot generate enough successful trajectories to learn from. AMD bypasses that by transferring three complementary memory types from a GPT-4-class teacher at inference time: workflow memory for task-level strategy, subtask memory for concrete behavioral examples, and a third layer for finer-grained cues. No fine-tuning is required on the student side. For teams deploying 7B-scale agents in production, this is a training-free way to close a meaningful capability gap without touching model weights. link

05 [Eval] SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring Nearly 60% of unsolved SWE-bench Verified instances contain flawed tests, making current leaderboard rankings unreliable as a signal of agent capability. SWE-Bench ProMax shifts the evaluation surface to large-scale multilingual code refactoring, a task that requires behavior-preserving changes across many files and cannot be gamed by memorizing gold patches from training data. The new benchmark is harder to saturate and harder to cheat. Teams using SWE-bench scores to make model selection decisions should treat those numbers with more skepticism than the leaderboard presentation implies. link

06 [Training] Omega-S: A Functional Resilience Index for LLM Fine-Tuning Catastrophic forgetting during LoRA fine-tuning can be reduced without storing old weights, computing Fisher matrices, or touching previous-task data. Omega-S derives a forgetting penalty directly from the weight matrix geometry, adds three lines to an existing training loop, and costs under 4% extra per step. On Llama-3-8B fine-tuned from code to prose, it lifts HumanEval pass@1 retention from 0.173 to 0.238 across 9 of 10 seeds. Any LoRA workflow that crosses domain boundaries at fine-tune time has an essentially free regularization option here. link