Also Worth Noting - 2026-08-23

Also Worth Noting

02 [RAG] GTA-RAG: Graph-Trajectory-Augmented Reinforcement Learning for Multi-Turn Retrieval-Augmented Reasoning Training RL agents on graph-structured retrieval trajectories, not flat search histories, gives multi-hop reasoning a reusable scaffold that cuts compounding errors past hop 3. Standard RL approaches reward only the final answer, which leaves the model blind to whether it actually assembled the right evidence chain along the way. GTA-RAG encodes each retrieval step as a node in a trajectory graph, giving the reward signal a structural view of the full evidence path. Teams running iterative RAG pipelines on multi-hop QA should watch whether trajectory-shaped supervision closes accuracy gaps that final-answer rewards cannot. link

03 [Eval] ExecRubrics: Executable Tool-Augmented Rubrics for Verifiable and Efficient Long-Form Evaluation Natural-language rubrics are silently wrong whenever scoring criteria interact, because they cannot express penalties, conditional overrides, or alternatives through a linear weighted sum. ExecRubrics encodes evaluation logic as verifiable Python scoring functions, making the rubric's intent inspectable and reproducible without a black-box LLM judge. Any response where one criterion should override or discount another exposes the flaw in current judge pipelines. Teams shipping LLM-judge infrastructure should audit whether their rubrics contain conditional logic that a weighted sum is quietly flattening. link

04 [Inference] SANE: State Anomaly Neutralization for Stable Extreme-Context Delta-Rule Models At 100 million tokens, RWKV-7 fails via localized norm explosion in a sparse subset of state channels, not the global saturation most practitioners would expect. Persistent decay keeps weakly updated entries small, but uneven injections let a few channels accumulate extreme values, destabilizing the fixed-size recurrent state. SANE identifies and neutralizes those outlier channels at inference time, with no fine-tuning required. For teams deploying recurrent models on very long contexts, this is a patch-at-inference fix rather than a retrain. link

05 [Agent] Small Reasoning Models are Instruction Followers in Function Calling Small reasoning models match or beat fine-tuned native function-calling models when given structured instruction-following prompts, which means the context format matters more than the fine-tuning. The IFFC framework decouples function-calling logic from the tool-calling context, routing it through standard user-assistant interactions instead. That framing alone unlocks accuracy gains without any additional SFT. Teams spending GPU budget on function-calling fine-tuning may be solving the wrong problem. link

06 [Training] Scaling Curriculum Learning For Autonomous Driving Scaling batched RL simulators to billions of driving interactions has not improved sample efficiency, because uniform domain randomization wastes most samples on already-mastered scenarios. Curriculum learning fixes this by adaptively prioritizing scenarios that sit at the current policy's learning frontier, cutting wasted simulation compute without changing the underlying RL algorithm. The diagnosis is structural: high throughput feeds the algorithm faster than it can learn when scenario difficulty is ignored. Teams running large-scale driving RL should treat scenario ordering as a first-class training hyperparameter. link