01Agent demo
A four-agent team with no assigned roles or game-specific logic completed Minecraft by writing and debugging 2,783 programs. It is a vivid stress test of coordination and tool use: the team won after 27 hours 14 minutes, but died 605 times on easy difficulty.
Why it mattersIt is a vivid stress test of coordination and tool use: the team won after 27 hours 14 minutes, but died 605 times on easy difficulty.
Keep in mind: The result is author-reported in one game.
Read the source · arXiv ↗02Reality check
Realtime-Bench evaluates dynamic tasks where delayed actions change the outcome under actual wall-clock inference latency. The best wall-clock success was only 44 percent and best Pass^3 was 36 percent; extra reasoning had mixed effects under time pressure.
Why it mattersThe best wall-clock success was only 44 percent and best Pass^3 was 36 percent; extra reasoning had mixed effects under time pressure.
Keep in mind: Coverage is limited to the benchmark's dynamic tasks.
Read the source · arXiv ↗03Agent engineering
A proposed onboarding agent jointly selects the model, prompt and harness from only tens of examples for a specialized workflow. The reported system matched or beat a strong baseline at up to 29 times lower cost, making agent configuration an engineering lever rather than a one-off craft exercise.
Why it mattersThe reported system matched or beat a strong baseline at up to 29 times lower cost, making agent configuration an engineering lever rather than a one-off craft exercise.
Keep in mind: The comparison is limited to the authors' workflow and cost assumptions.
Read the source · arXiv ↗04AI for science
PhysClosureBench tests fifty expert-verified physics investigations and scores both local accomplishments and complete research closure. The top model averaged 6.29 of 10 but reached full closure in only 32.7 percent of attempts; 36.2 percent of attempts scoring at least 8 still remained incomplete.
Why it mattersThe top model averaged 6.29 of 10 but reached full closure in only 32.7 percent of attempts; 36.2 percent of attempts scoring at least 8 still remained incomplete.
Keep in mind: The benchmark covers a specialized set of physics tasks.
Read the source · arXiv ↗05Efficient agents
The authors replace selected generative routing and review steps with a lightweight decision model in multi-agent pipelines. The approach cut generative tokens by 12.7 to 54.6 percent while staying within 1.34 accuracy points and lowering latency in seven of eight settings.
Why it mattersThe approach cut generative tokens by 12.7 to 54.6 percent while staying within 1.34 accuracy points and lowering latency in seven of eight settings.
Keep in mind: Results are author-reported and benchmark-dependent.
Read the source · arXiv ↗06Agent safety
AgentPerturb systematically alters agent interactions and traces what happens after corrupted observations or tool outputs enter a trajectory. Agents rarely recognized the altered inputs, propagated corrupted information without verification, and successful perturbed paths could require substantially more computation.
Why it mattersAgents rarely recognized the altered inputs, propagated corrupted information without verification, and successful perturbed paths could require substantially more computation.
Keep in mind: The paper says code and data will be released, so reproducibility is pending.
Read the source · arXiv ↗07Collective reasoning
A test-time framework normalizes reasoning styles, then deliberately places an agent's preferred answer in the minority to expose unstable decisions. Across three mathematical reasoning benchmarks and four heterogeneous model combinations, the authors report higher average accuracy and better token efficiency than competing non-oracle aggregators.
Why it mattersAcross three mathematical reasoning benchmarks and four heterogeneous model combinations, the authors report higher average accuracy and better token efficiency than competing non-oracle aggregators.
Keep in mind: The abstract does not report exact aggregate accuracy gains.
Read the source · arXiv ↗