The 2026-11-06 YZ Index Smoke quick test covered 13 models, with Doubao Pro ranking first for the day at 87.6 points. Because the daily sample is small, the results are best viewed as short-term monitoring signals rather than long-term conclusions about model capability.
Deployed AI systems can behave very differently from their foundation models, so safety scores from system cards cannot be assumed to transfer. Teams should test complete deployment systems and account for measurement uncertainty.
Developers and deployers can now use two MLCommons benchmarks, AILuminate Safety 1.2.1 and AILuminate Jailbreak 1.1, to test AI systems' propensity to produce harmful content and how easily attackers can induce violations. Both benchmarks can be run on demand through MLCommons' new Candela service, and new public results for 27 systems have been released.
MLCommons has launched Candela, a benchmarking-as-a-service offering built on the AILuminate benchmarks to measure AI risk and reliability. It enables organizations to test complete AI systems across configurations and cite results for governance and procurement.
On this page1. CP × HiCache: Page-based KV sharding1.1 Logical radix tree and page ownership1.2 Reading sharded KV: Gather one layer at a time1.3 HiCache integration: Restore each shard before gatheri
On this pageTL;DRHow Scale-Up Builds on RecoveryOne Logical Deployment, Three Cooperating LayersHow a Live EP Group GrowsMeasuring the TransitionRoadmapEPLB after scale-upCUDA graph recaptureMulti-ran
On this pageBuilt on TorchNeuron from the startWhat's in the betaEarly preview: heterogeneous RLJoin the betaOver the past year, we have continued to expand the hardware SGLang supports, alongside lau
On this pageTL;DRBackgroundWhy agentic RL is differentArchitectureThe stackExperimental setupWhere the time goesStep time breakdownThe long-tail problemContext growth and the KV cacheThe CPU side: san
On this pageWhat is integratedResultsWhere the speed comes fromWhat SGLang addsAcknowledgmentsAppendix: benchmark setupPrefillDecodeTL;DR On par or ahead on DeepInfer’s own benchmarks. Up to 1.04x pre
On this pageTL;DRAbstract1. Introduction2. Why Native NVFP4 RL Becomes Unstable2.1 Failure Develops During Training2.2 Quantization Error, Conversion Error, and Execution Mismatch2.3 Activation Precis
On this pageImpact and affected deploymentsWhat happenedImmediate remediation and upstream fixRelated KV cache lifetime issuesBackgroundThe Problem: When Is a KVCache Slot Actually Free?Impact and Exp
The 2026-11-05 YZ Index Smoke quick test covered 13 models, with GPT-6 Astra and GPT-6.1 Sol tied for first at 81.66 points. The Smoke test is a daily 10-question quick test suited to short-term signals and is not equivalent to Full weekly leaderboard conclusions.
In the latest WDCD v3.1 round, Claude Opus 4.7 rose 8 points to 89.30 while Doubao Pro dropped 10.1 points, producing a pattern of three gains and five declines that sharply exposes differences in how stably models hold to constraints under multi-round pressure.
In WDCD v3.1, business rules are the hardest of the five major constraints to adhere to, with doubao-pro scoring only 1.2/4, while data boundaries show the widest differentiation and GPT-5.5 a 2.7-point cross-scenario gap.
Across 15 models tested on eight v2 anchor prompts over three rounds, average R1 confirmation reached 0.93, average R2 resistance fell to 0.67, and average R3 integrity held at 86.7%, with Qwen3 Max the only full collapse. The other 14 models recorded no R3 collapse.
In the WDCD v3.1 commitment-keeping test, GPT-6 Astra ranked first with 90.50 points while Doubao Pro finished last at 58.70, with scores across the 15 evaluated models spanning 90.50 to 58.70 — a 31.8-point range.
In today's Smoke evaluation, Claude Sonnet 4.6's main leaderboard score fell from 84.90 to 73.76, a drop of 11.1 points, driven by double-digit declines in both code execution and material constraint. Its integrity rating also shifted from pass to warn.
In today's Smoke evaluation, GPT-6 Luna's main leaderboard score dropped 11.5 points from 87.27 to 75.79, driven by a 25.7-point collapse in the code execution dimension from 100.00 to 74.30, while its integrity rating moved from pass to warn.
The 2026-11-04 YZ Index Smoke Quick Test covered 12 models, with GPT-5.5 ranking first for the day at 94.8. The daily 10-question test is a short-term signal and is not equivalent to the weekly Full ranking.
On 2026-11-03, the YZ Index Smoke quick test covered 13 models, with GPT-5.5, GPT-6 Astra, GPT-6 Luna, and GPT-6 Sol tied for first at 87.27. The Smoke test uses 10 questions per day and is suitable for observing short-term signals, not for drawing Full weekly leaderboard conclusions.
