Baidu Wenxin Discloses ERNIE Evolve RSI System, Used in Wenxin 5.5 Flash Inference Optimization
On November 9, 2026, Baidu Wenxin disclosed its recursive self-improvement (RSI) progress, unveiling the ERNIE Evolve auto-research system used in developing the upcoming Wenxin 5.5 Flash. ERNIE Evolve achieved state-of-the-art results on three benchmarks across NanoChat AutoResearch and NanoGPT Speedrun tasks, and has been applied to optimize inference for Kunlunxin P800, including low-latency performance, deterministic throughput, and decoding speed. The system employs a dual-loop architecture with multiple agents (Director, Selector, Executor, Evaluator, Reviewer) to autonomously propose research directions, run experiments, and retain only verified improvements.
Coverage timeline
量子位量子位
百度文心首次披露RSI进展:ERNIE Evolve已用于最新模型研发,文心5.5 Flash 或将近期发布 量子位的朋友们 2026-11-09 17:48:04 来源: 量子位 11月9日,百度文心首次披露 RSI 最新进展 11月9日,百度文心首次披露 RSI 最新进展,公布参与文心 5.5 Flash 研发的自动研究系统 ERNIE Evolve,其能够围绕目标与约束,自主提出研究方向、开展实验、验证结果并积累经验,既可以持续优化已有路径,也能根据实际效果探索新方向。在 NanoChat AutoResearch 和 NanoGPT Speedrun 两项任务的三个评测指标上,均取得了目前公开验证的 SOTA 表现。目前,ERNIE Evolve 已应用于即将发布的文心 5.5 Flash 推理优化,包括昆仑芯 P800 低时延场景的性能优化、确定性推理吞吐优化及极致解码速度优化。 ERNIE Evolve 基准测试表现 递归自我改进(Recursive Self-Improvement,简称 RSI)是当前全球 AI 领域的重要探索方向之一,核心理念是让 AI 直接参与自身研发流程,不断优化自身能力,形成自我改进闭环。OpenAI、Anthropic、腾讯、智谱等多家厂商都在探索将 RSI 引入模型研发。 ERNIE Evolve 采用双层循环组织研究,外层循环结合历史实验结果和已有知识,提出新的研究方向并安排多条研究线并行探索;内层循环则针对具体方向持续进行代码修改、效果评测和经验沉淀。系统内部由多个智能体分工协作,Director 负责选择研究方向,Selector 负责确定每轮实验的目标和代码起点,Executor 负责修改代码并生成新版本。每轮实验结束后,独立评测程序 Evaluator 负责评测代码性能和实验结果,Reviewer 则进一步检查研究是否偏离目标、是否存在评测漏洞。只有通过评测和复核的改进才能保留。每轮实验的代码改动、结果和经验都会被记录,供后续研究参考。 ERNIE Evolve 架构 将 ERNIE Evolve 接入文心 5.5 Flash 真实推理优化后,系统在生产环境中持续取得显著收益。 在 昆仑芯 P800 低时延场景优化 中,ERNIE Evolve 历时约 168 小时,经过 4 轮方向探索,先后接入原生融合算子并启

机器之心新闻资讯
11月9日,百度文心首次披露 RSI 最新进展,公布参与文心 5.5 Flash 研发的自动研究系统 ERNIE Evolve,其能够围绕目标与约束,自主提出研究方向、开展实验、验证结果并积累经验,既可以持续优化已有路径,也能根据实际效果探索新方向。在 NanoChat AutoResearch 和 NanoGPT Speedrun 两项任务的三个评测指标上,均取得了目前公开验证的 SOTA 表现。目前,ERNIE Evolve 已应用于即将发布的文心 5.5 Flash 推理优化,包括昆仑芯 P800 低时延场景的性能优化、确定性推理吞吐优化及极致解码速度优化。 ERNIE Evolve 基准测试表现 递归自我改进(Recursive Self-Improvement,简称 RSI)是当前全球 AI 领域的重要探索方向之一,核心理念是让 AI 直接参与自身研发流程,不断优化自身能力,形成自我改进闭环。OpenAI、Anthropic、腾讯、智谱等多家厂商都在探索将 RSI 引入模型研发。 ERNIE Evolve 采用双层循环组织研究,外层循环结合历史实验结果和已有知识,提出新的研究方向并安排多条研究线并行探索;内层循环则针对具体方向持续进行代码修改、效果评测和经验沉淀。系统内部由多个智能体分工协作,Director 负责选择研究方向,Selector 负责确定每轮实验的目标和代码起点,Executor 负责修改代码并生成新版本。每轮实验结束后,独立评测程序 Evaluator 负责评测代码性能和实验结果,Reviewer 则进一步检查研究是否偏离目标、是否存在评测漏洞。只有通过评测和复核的改进才能保留。每轮实验的代码改动、结果和经验都会被记录,供后续研究参考。 ERNIE Evolve 架构 将 ERNIE Evolve 接入文心 5.5 Flash 真实推理优化后,系统在生产环境中持续取得显著收益。 在 昆仑芯 P800 低时延场景优化 中,ERNIE Evolve 历时约 168 小时,经过 4 轮方向探索,先后接入原生融合算子并启用投机解码,再对主机调度与同步机制、通信路径进行优化,最优版本的生成速度提升至基线的 623.6%。 在 确定性推理吞吐优化 中,ERNIE Evolve 用 3 天时间围绕确定性采样、确定性开关及算子并行策略开展 3 个方向的探

Related stories
Baidu open-sources ERNIE 5.5 Flash Base, 280B MoE model with 14B active parameters
On November 5, Baidu open-sourced ERNIE 5.5 Flash Base, a sparse MoE language model with 280B total parameters and 14B active per token, pretrained on 33.5T tokens with up to 128K training context and 1M-token context for fine-tuning and inference. The company claims SOTA results among same-size models on benchmarks covering knowledge, reasoning, math, code, multilingual, and long-context tasks, surpassing DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base. The release includes weights on Hugging Face and ecosystem support with vLLM and SGLang.
3 reports · 2 sources

LoopVL open-source VLM shows data efficiency with 260x less training data
Researchers from Renmin University, Baidu, and others released LoopVL, an open-source vision-language model based on the Loop Transformer. Despite using only 0.14T training tokens versus about 36T for Qwen3.5-2B and InternVL3.5-2B, LoopVL matches or exceeds them on benchmarks like LogicVista, MMMU-Pro, and MathVision, demonstrating high data efficiency. The team also observed cross-loop visual attention reorganization, termed 'Visual Aha Moment'.
1 reports · 1 sources

Google launches Gemini Omni 1.2 Flash with 15-second clips, 30% lower cost
Google has rolled out Gemini Omni 1.2 Flash, a video-generation model that creates clips up to 15 seconds, compared to 10 seconds with Omni 1.1. The new model improves reference adherence and costs 30% less than its predecessor. It is being distributed through the Gemini API and Google AI Studio.
2 reports · 2 sources
DeepSeek releases V4.1-Flash-1017 open-weight model with faster, cheaper agent tasks
DeepSeek released V4.1-Flash-1017, an open-weight model checkpoint with 763B parameters, now live on its website, app, and API. The update claims stronger benchmark performance, faster task completion, and shorter outputs than its predecessor, achieved via reinforcement learning with self-feedback for end-to-end project development. DeepSeek also opened its Search & Fetch APIs to everyone, priced at $0.75 per 1000 successful requests, and requires DeepSeek Harness v0.2.1-rc.1 or later for optimal use.
9 reports · 6 sources955 points · 717 comments

News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe