Shopee’s CompassMax-V3-Thinking paper shows that scaling RL for giant MoE models is less about buying more rollouts and more about making every rollout produce usable learning signal.
A mechanism-first reading of TROJail, showing why multi-turn jailbreak risk is less about one bad prompt than about trajectory-level strategy, sparse credit assignment, and semantic drift.
SymPyBench shows why scientific AI evaluation needs executable ground truth, controlled variants, and robustness metrics beyond headline accuracy.
A close reading of how a GPT-5-based correctness checker turns scientific paper auditing from artisanal peer-review labor into a scalable quality-control workflow.
A comparison-based reading of how a multimodal oncology agent turns generated clinical reports into measurable predictive signal for IDH1 mutation status in low-grade glioma.
A mechanism-first reading of VQR-DQN, showing where quantum feature extraction may help resource-allocation RL—and where the evidence still stops.
PRiSM shows why high final-answer accuracy is not enough for multimodal scientific reasoning, and how businesses should evaluate AI systems that must handle diagrams, formulas, code, and uncertainty.
A case-first look at how a multi-step LLM pipeline converts therapy transcripts into clinician-verifiable personalized networks, and why that matters more than another clever summary bot.
TRACE shows how vision-language model evaluation can move from final-answer scoring to step-level diagnosis, confidence triage, and failure localization.
GraphBench shows why graph learning needs broader, harder, and more realistic evaluation before anyone should trust claims about general-purpose graph intelligence.