Applied to the Qwen 3.8 2.4T model on NL2RepoBench, QwenGyre increased accuracy from 52.5% to 58.5% in 48 training steps, completing the steps in 75.42 hours compared to 134.55 hours for asynchronous training.

Overview Debates Related claims Topics Sources

Artificial Analysis reported that MiniMax M3.1 Flash demonstrated a 93 percent hallucination rate on the AA-Omniscience benchmark because the model attempted nearly every question rather than acknowledging uncertainty.

Artificial Analysis found MiniMax M3.1 Flash's Intelligence Index task evaluation cost dropped 57 percent to $0.22 from MiniMax M3's $0.51, compared to $0.26 for DeepSeek V4.1 Flash-1017, $0.28 for GLM-5.3 Flash, and $1.24 for Gemini 3.8 Flash.

MiniMax M3.1 Flash achieved an overall AA-Omniscience benchmark score of 16 despite correctly answering 56 percent of factual questions, compared to 43 percent for DeepSeek V4.1 Flash-1017 and 28 percent for GLM-5.3 Flash.

Artificial Analysis awarded MiniMax M3.1 Flash a score of 42 on its Intelligence Index, an increase of 13 points over the MiniMax M3 model's score of 29.

Artificial Analysis reported that DeepSeek V4.1 Flash-1017 scored 45 on its Intelligence Index, while Gemini 3.8 Flash scored 41.

MiniMax reported that all five reasoning effort levels for MiniMax M3.1 Flash stayed below $0.05 per task on its Automation-Bench chart.

Artificial Analysis reported that DeepSeek V4.1 Flash-1017 generated approximately 224 million output tokens during its Intelligence Index evaluation.

Artificial Analysis found MiniMax M3.1 Flash generated approximately 189 million output tokens during its Intelligence Index evaluation, compared to roughly 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash.

Frenos emphasized that its cyberattack tests used supplied information and simulated scenarios without proving that artificial intelligence models could independently identify targets or execute cyberattacks.

The Frenos study evaluated 10 artificial intelligence models across 16,980 total tests conducted between October 2025 and September 2026, including 13,884 decision attempts, 2,520 planning attempts, and 576 simulated attack runs.

The three leading frontier models tested by Frenos produced attack plans that passed every repeat test in 60%, 62.9%, and 74.3% of attempts, showing inconsistency and omitting essential steps or assuming unestablished access.

Single decision accuracy varied across models in the Frenos study, with models correctly matching required outcomes in 6% to 47% of single action choices.

Anthropic technical staff researcher Justin Curl stated that defining quality in the legal domain is notoriously difficult for artificial intelligence research

Anthropic offers four-month Conceptual Reasoning Fellowships paying between $100,000 and $167,000 to recruit legal experts to create evaluation test cases for its Claude AI models

DeepSeek trained DeepSeek-V4.1-Flash-1017 using reinforcement learning with self-feedback to improve performance on end-to-end coding, game development, and 3D modeling tasks

DeepSeek's reinforcement learning with self-feedback method uses internal confidence scores and self-generated critiques as reward signals, eliminating reliance on human evaluators or external reward models

Reflection AI released post-training recipes, code templates, and safety measurement tools alongside Beam's weights on October 16, 2026, to enable developers to fine-tune the Beam model

Reflection AI trained Beam using intensive reinforcement learning that penalized the model whenever it generated unnecessary tokens to complete a task

Reflection AI developed automated reliability optimization algorithms that detect data quality errors and revert training runs to earlier stages instead of restarting from scratch

A person close to Tencent stated that the scraping fleet originated from tests by a Tencent evaluation team checking how third-party models browse the web and answer questions.