Loading the edition…
· Past edition · 3 stories
Retrospective edition reconstructed from later captures. Paper versions may include later updates; coverage labels describe the evidence used.
Researchers report retrosynthesis results, while engineering teams explore agent evaluation and multi-GPU serving.
Applications & products
RetroChimera combines two retrosynthesis models with complementary strengths, then uses learned rank-dependent voting to rank their proposals.
Why it matters: If its proposed routes help chemists assess more candidate molecules, synthesis planning could become more efficient. The source presents this as a potential benefit, not a guarantee of experimental success.
Agent frameworks & orchestration
Open each story in the reader. Summaries are shown when available; some sources provide an abstract only. With an account you can keep your own notes.
A valid tool call is not a completed task: agent evaluation must check both the steps in a trace and the environment’s final state.
Why it matters: A model can reach the right outcome despite redundant calls, or make accurate calls while failing the task. Separate outcome and process scores show what users experience and where to debug, tune, or reduce wasted steps and cost.
Infrastructure & deployment
Dynamo-Triton 26.07 lets one gRPC model endpoint serve a compiled TensorRT plan across multiple GPUs, keeping rank coordination out of the client.
Why it matters: This offers teams a way to trade more GPUs for shorter request latency without making the client coordinate ranks. The benchmark supports a latency case, not a general cost or throughput conclusion: concurrent throughput, cost per video, and TCO were not measured.