Baidu announced ERNIE Evolve on Monday, showing an agentic system that edits code and runs tests to optimize machine learning software. The Next Web reporter Ana Maria Constantin reported that the tool lifted inference speeds for Baidu’s unreleased ERNIE-5.5-Flash model [1]. On social network X, Baidu framed the project as a real step toward recursive self-improvement (RSI), where AI tools tune the code that trains and serves neural nets. Self-reported data shows the system boosted speeds on Kunlunxin P800 chips and NVIDIA GPUs, though outside groups stress that all tests remain closed [1, 2].
How ERNIE Evolve Runs Experiments
Given a code repo, a target goal, and test rules, ERNIE Evolve runs two nested loops to find valid code gains. The outer loop picks what to test across parallel research lanes, while the inner loop builds and checks specific code edits within each chosen path. A planning agent named the Director sets the next route and picks a base code version [2].
To pick smart paths, Director reads an Experience Bank of past test logs and a Knowledge Base of papers and code. Parallel lanes build trees of code versions so hypotheses stay apart. Failed tests stay in the log. That log keeps later rounds from repeating dead ends [2].
Inside each lane, a Selector picks a branch from the code tree and names a hypothesis to test. An Executor then writes the patch inside an isolated sandbox so unfinished code can’t break live systems. Two separate guards check the work. An Evaluator runs as a fixed program kept outside the repo where no agent can see it, and code gains must top a noise limit to pass. Next, a Reviewer inspects the code to check that it solves the real task without cheating the test rules. Edits stay only when they pass both checks. MarkTechPost reported that this dual check stops agents from gaming internal scores [2].
Gains on ERNIE-5.5-Flash
Baidu tested the system on three live tasks for ERNIE-5.5-Flash. On Kunlunxin P800 chips, made by the Baidu chip unit eyeing an IPO, text generation speed jumped to 623.6% of baseline across four research rounds [1, 2]. Round 1 added fused mHC operations and speculative decoding through the multi-token prediction layer to hit 282.9% of baseline. Round 2 cut host-side overhead to 588.8%. Round 4 fixed dispatch routines so the DeepEP low-latency path engaged, but China AI Dispatch noted that much of this surge came from turning on unused hardware features not just fresh code tricks [1, 2]. The Kunlunxin test was not a direct match for GPU runs [1].
On graphics cards, the agent tackled deterministic inference for reinforcement learning under SGLang mode, raising throughput by 33.7% in about three days in Baidu tests covered by The Next Web [1, 2]. Time per output token fell 26%. Time to first token dropped 12%. A Triton kernel for deterministic sampling delivered the largest single gain after previously taking about 14% of GPU run time [2].
A third test targeted single-request decoding on GPUs. Three out of seven research directions yielded real gains, pushing single-request decode speed up by 14.9% across live benchmarks. Key changes included FP16 instructions for tiny FP8 matmuls and a two-level MoE expert scheduler. Baidu did not name the exact GPU model used in these production tests, though company engineers had already tuned the base software before the agent started its work [2].
Public Benchmark Results
Baidu also tested ERNIE Evolve on public AI research challenges using single NVIDIA B200 cards and clusters of eight NVIDIA H100 GPUs. The results were mixed. On NanoChat AutoResearch, the agent scored 0.8763 validation bits per byte (BPB) within a five-minute budget on one B200. It hit that mark by combining the base network with text retrieval tools and n-gram lookups, routing token flows through a learned gate. That score beat rival systems such as Tencent Hunyuan’s Hyra-1.0 at 0.9015 BPB, (auto)²-research from rekursiv.ai at 0.8878 BPB, and Recursive at 0.9109 BPB [2].
On NanoGPT Speedrun Track 1, the system hit target loss in 38.9 seconds on eight NVIDIA H100 GPUs, edging past the 39.9-second public record [1, 2]. Baidu reported 38.92 seconds across eighteen runs. On Track 3 of that challenge, the agent trimmed training from the 2,690-step mark down to 2,590 steps by splitting weights into direction and magnitude while adding EMA-Nesterov extrapolation [2]. But China AI Dispatch noted that Baidu has not submitted that run to the official leaderboard, leaving the record unverified by independent organizers [1].
Baidu did not sweep every test. On SOL-ExecBench, which evaluates kernel optimization across 235 tasks on an NVIDIA B200, ERNIE Evolve posted a mean score of 0.819. That lagged behind the leaderboard leader at 0.827. Still, it finished ahead of Tencent’s Hyra-1.0 at 0.754 and Recursive at 0.676 under Baidu’s eval stack. However, MarkTechPost pointed out that Recursive and Hyra-1.0 had set their older times against a loose ~79.7-second Speedrun baseline, so those figures are not direct head-to-head matches [2].
Industry Context for ERNIE Evolve
The race to get AI systems to improve other AI systems is heating up across global technology companies, as reported by Ana Maria Constantin for The Next Web. Anthropic says Claude now writes over 80% of its own code. At Google, co-founder Sergey Brin confirmed that Gemini assists in developing its successor. Meanwhile, AI startup Recursive claimed its internal agent system largely built the Flint coding model. Baidu’s published benchmark table claims ERNIE Evolve beats Recursive across Speedrun and NanoChat tests [1, 2].
While engineering teams elsewhere explore specialized architectures like Liquid AI models for on-device AI agents, Baidu built its agent to tune server code for large foundation models. Liquid AI focuses on lean, edge-ready neural nets. In contrast, Baidu targets data center efficiency, tuning low-level kernels on server clusters without changing model weights.
Despite sharing benchmark scores and code logs on GitHub, Baidu has kept the inner ERNIE Evolve agent closed to the public. The company has not revealed which large language models power the Director, Selector, or Reviewer agents. Outside coders cannot download the agent loop. They can inspect code diffs and GitHub logs, but Baidu keeps its agent prompts and orchestration setup private [2].
Supervision and Future Plans
Baidu next plans to expand ERNIE Evolve from inference tuning to model training. Later, it aims to let the agent design its own evaluation tests [1]. But letting AI build its own tests raises the danger of reward hacking, where an agent exploits blind spots to post high scores without real software gains. Baidu’s hidden Evaluator and separate Reviewer were built specifically to catch those tricks before bad code enters production [2].
Baidu stated that recursive self-improvement must stay under human control. “Its scope should grow only while people can still understand, supervise, and stop the system,” Baidu wrote in its release. Engineers keep the power to halt runs at any moment. For now, ERNIE Evolve acts as a smart assistant for Baidu coders, not an independent agent left to run on its own [1].
- ONLINE NEWS Constantin, A. M. (2026, November 9). Baidu says its AI research agent sped up its next ERNIE model. The Next Web. [Article Link]
- ONLINE NEWS MarkTechPost. (2026, November 9). Baidu’s ERNIE Evolve Automated Research System Beats NanoChat and NanoGPT Records. MarkTechPost. [Article Link]
