Also Worth Noting - 2026-07-30

Also Worth Noting

02 [Training] CADENCE: Closing the Reasoning Gap via Coverage-Adaptive On-Policy Distillation Cold-start collapse can kill a distillation run before the student learns a single useful token. CADENCE identifies three compounding failure modes in on-policy knowledge distillation and fixes each within a unified framework: a DRIFT mechanism adapts KL scheduling to the student's actual coverage state rather than a time-only schedule, and a graded reward signal replaces binary pass/fail to recover information from partially correct traces. Treating all three as a joint problem rather than separate patches is what separates this from prior work. Teams distilling large reasoning models into compact ones should audit which of these three failure modes is silently degrading their runs first. link

03 [Inference] TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM Running a vision-language-action model at 32 Hz on a single RTX 4090 with under 1 GB VRAM is not an optimization trick applied to the standard architecture. TurboVLA replaces the conventional V→L→A pathway, where visual observations route through a large language model before reaching the action decoder, with a direct V+L→A mapping that keeps vision and language processing independent from the action head. Bypassing the LLM-centric bottleneck collapses memory by more than an order of magnitude compared to standard VLA designs. Robotics teams blocked by GPU cost now have a documented path to capable real-time policies on consumer hardware. link

04 [Eval] SecRespond: Benchmarking AI Agents for Real-World Post-Compromise Incident Response Every existing cybersecurity agent benchmark tests a clean pre-attack environment, which means the entire post-compromise workflow has gone unmeasured. SecRespond fills that gap as the first benchmark to evaluate LLM agents handed a compromised host with live artifacts and a command-line interface, mirroring how security operations teams actually encounter incidents. The pre-compromise focus in prior work is not a minor omission; it means current leaderboard numbers say nothing about whether an agent can triage a real breach. Security teams evaluating LLM agents for SOC workflows should treat pre-compromise benchmark scores as a separate, largely irrelevant signal. link

05 [Open-source] GPT-Red: Automated Red Teaming via Self-Play at Scale OpenAI used a self-play red-teaming agent to adversarially train GPT-5.6, and the methodology is now publicly documented. GPT-Red attacks a population of simultaneously trained defender agents rather than a fixed target, forcing the red-teamer and defender to co-evolve at the same compute scale used for frontier training runs. The paper reports specific attack success rates before and after adversarial training, giving teams a concrete before/after baseline to reference. Teams building prompt-injection defenses can treat this as a replicable hardening recipe rather than an opaque capability claim. link

06 [Agent] SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution Standard agentic RL discards cross-task structure entirely, treating each episode as independent even when tasks share reusable solution patterns. SkillRise organizes related task instances into progressively challenging sequences and runs a single policy that alternates between task execution and skill evolution within one RL loop, avoiding the multi-stage pipelines that entangle extraction, retrieval, and execution into separate brittle components. The unified objective outperforms both single-task RL and pipeline-based skill reuse on multi-task agent benchmarks. Teams running RL for multi-task agents should look at whether their current setup is throwing away transferable structure that costs nothing to keep. link