Also Worth Noting - 2026-08-28

Also Worth Noting

02 [Agent] What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents Generating more agentic interaction data does not guarantee better agents. A new two-level framework reframes quality around three axes: consistency across environments and tasks, coverage of meaningful scenarios, and executability of success signals. Most current pipelines conflate candidate construction with verification, which inflates volume while leaving transfer gaps. Teams auditing their agentic data pipelines now have a concrete three-axis checklist to find where generated experience will fail to transfer before training begins. link

03 [Training] TTPO: Test-Time Policy Optimization Majority-vote pseudo-labels can replace ground-truth labels for test-time policy optimization, but the failure mode is asymmetric: a wrong vote corrupts the teacher and misleads every subsequent token, not just the contested one. TTPO fixes this by treating disagreeing rollouts as a reliable negative signal regardless of whether the vote itself is correct, breaking the corruption chain without requiring any labeled data. This opens RL-style post-training to domain-specific deployments where labeled ground truth simply does not exist. link

04 [Theory] Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO Evolution Strategies cover a broader slice of the reasoning distribution than GRPO, and the mechanism is structural rather than incidental. GRPO collapses toward a single reference model, which prunes low-probability but correct reasoning paths early; ES avoids that collapse by maintaining a wider policy distribution throughout training. The advantage is most pronounced when the reward surface is sparse and GRPO mode-collapses before the model finds valid solutions. Teams hitting GRPO ceilings on hard reasoning tasks should treat ES as a concrete alternative, not a theoretical curiosity. link

05 [Hardware] TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback Chunk-based vision-language-action models are blind to contact changes for up to 500ms because tactile conditioning goes stale the moment a chunk starts executing. TacForcing streams tactile feedback mid-chunk rather than conditioning only at chunk boundaries, keeping the action signal current throughout execution. On peg-insertion tasks, this mid-chunk update cuts failure rates by a reported 40% without requiring a separate high-frequency controller. For teams deploying VLA models on contact-rich manipulation, latency in the feedback loop is now the variable worth measuring first. link

06 [Eval] PAWBench: How Far Are We from Probabilistically Aligned World Modeling? Current video world models are evaluated on whether a single generated trajectory looks plausible, not on whether repeated generations reproduce the correct distribution of physically valid outcomes. PAWBench introduces probabilistic alignment as a distinct evaluation axis: a model that always produces one correct trajectory still fails if it cannot recover the full distribution of valid behaviors from the same initial state. No existing benchmark measures this, and the gap between single-trajectory plausibility scores and distributional coverage turns out to be substantial. Teams framing video generators as world models for planning or simulation should treat distributional coverage as a first-class metric. link