DeepSeek-V4.1-Flash tops open-source Code Arena as agents negotiate alone

The shift to perfect-document scoring changes how extraction systems should be evaluated: one wrong cell fails the whole doc, which mirrors the economics of enterprise workflows where partial accuracy is often unusable. Teams building document pipelines can use the benchmark and its synthetic template pipeline to stress-test models before deployment without exposing private data. For comparison, its hardest cases sit well above what saturated benchmarks like ExtractBench already measure.

Jerry Liu@jerryjliu0Source

Hot-swapping weights into a running deployment removes the redeploy bottleneck that dominates wall-clock time in RL post-training, so teams can iterate on checkpoints far more cheaply. Engineers running RL pipelines or serving infrastructure should evaluate it against their current checkpoint-deploy scripts. The Nemotron 3.5 Lightning result doubles as a concrete case of using it to improve a search agent's benchmark score.

CoreWeave@CoreWeaveSource

The claim that models plan multiple tokens ahead challenges the token-by-token intuition most engineers carry, and "grown, not built" reframes how we should reason about model behaviour. Practitioners building on Claude gain a mental model for predicting failure modes and interpreting outputs. The discussion ties interpretability findings directly to alignment questions, useful for anyone weighing safety alongside capability.

Lenny Rachitsky@lennysanSource

The non-obvious move here is replacing score-based feedback with same-task A/B behavioural comparisons, which sidesteps the missing-verifier problem that blocks most self-improvement systems. Engineers building agent scaffolds can study the Propose–Probe–Compose loop and reuse its evidence-gated lesson library design in their own stacks. As an open-source release from Google Research, it is directly comparable to other RSI and agent-tuning frameworks that depend on task-specific verifiers.

shao__meng@shao__mengSource

Moving from residue-level to atomic-level generation lets researchers specify hydrogen bonds and exact atom placements, producing enzymes that actually catalyse the intended reactions. Protein engineers, drug designers and synthetic biology labs can now build with atomic precision using code that is already open source. Compared with RFdiffusion2, which works at the amino-acid level, RFdiffusion3 wins 37 of 41 active-site design tests.

The best single-game performer was not the most efficient learner, so one-shot benchmark scores can mislead when choosing an agent for long-horizon work. Teams building self-improving systems should track held-out learning curves against compute cost and verify that saved tools and notes are actually reused. The frozen-weights design makes the results directly relevant to anyone running agents under a fixed budget.

rohanpaul_ai@rohanpaul_aiSource

Percy Liang's closing argument reframes the KV cache as the structural root of Transformer inference inefficiency, pointing to architectures designed for inference from day one, such as SSMs, linear attention, and diffusion-based non-autoregressive generation. Engineers building or tuning inference stacks get a single coherent map from architecture to serving systems. The executable lecture notes with sympy-verified formulas make the material directly reproducible rather than slide-level theory.

shao__meng@shao__mengSource

Zero-training inference on tables removes the usual feature-engineering and fine-tuning cycle, which could make tabular ML accessible to teams without dedicated data scientists. Developers working with single tables or multi-table relational schemas can plug the library into existing PyTorch pipelines and get predictions from pretrained models directly.

官方一手marktechpost@Sana HassanSource