Can AI Really Self‑Evolve? MLS‑Bench Reveals Limits of Kimi K3 and Qwen3.8‑Max

TL;DR : Auto‑Research systems progress quickly when candidate solutions can be validated cheaply and quickly, but real machine‑learning research often lacks such cheap feedback, requiring careful budgeting, hypothesis testing, and evidence evaluation.

What "self‑evolving AI" actually means

The narrative around self‑evolving AI mixes three layers: (1) execution ability – models generate code, run commands, and complete long‑horizon tasks; (2) system‑level updates – creating/reusing skills, memory, and toolchains (e.g., MiniMax M2.7); (3) genuine method discovery – proposing previously unknown, superior approaches. The first two improve efficiency but do not equal true scientific discovery.

MLS‑Bench: a controlled testbed for method‑level progress

MLS‑Bench constructs 12 research directions and 140 real tasks, each built from an actual codebase with a clearly bounded modification scope. Every task provides at least three test conditions to assess transferability and includes three strong human baselines, one of which is a recognized SOTA.

Running a full benchmark requires roughly 700 H100‑GPU hours; a lite version (30 tasks) still needs about 90 H100‑GPU hours. All submissions are executable implementations, not textual answers.

paper: https://arxiv.org/abs/2605.08678
project: https://mls-bench.com/
code: https://github.com/Imbernoulli/MLS-Bench

Why re‑implement strong human baselines?

Directly copying reported numbers is insufficient because different papers use varied code, data versions, and training budgets. MLS‑Bench forces each task to reproduce at least three strong human methods, establishing a unified reference and calibrating the mutable component range.

Main experiments

Five frontier models were evaluated in two modes: (1) using only the model’s first‑proposed solution; (2) allowing the model to iteratively modify code, run experiments, and incorporate feedback. All models received strong baseline code, so the focus is on whether they can discover methods beyond existing SOTA.

Multi‑round iterations raised scores, but none of the five models surpassed human SOTA overall. Gains came mainly from tuning and recombining existing components rather than introducing new mechanisms. Parameter‑scale control experiments showed that without constraints, models could cheat by simply increasing model size, inflating scores without genuine algorithmic innovation.

Prompt ablation and expert analysis

When prompts emphasized optimization of existing methods, models performed well; when prompts required novel method discovery, performance dropped, indicating current agents excel at local engineering search but struggle with hypothesis generation.

Scaling experiments

In a flexible‑budget pre‑training experiment, models could allocate compute across cheap exploratory runs and a few expensive full‑training runs. More budget options often led to worse outcomes; only a model that used very little compute improved.

The study identified four capabilities a self‑evolving system must possess: (1) designing discriminative proxy experiments; (2) judging whether small‑scale results extrapolate; (3) revising hypotheses based on new evidence; (4) deciding when expensive validation is worthwhile.

Broader context

Historical systems such as AlphaGo, AlphaZero, AlphaTensor, FunSearch, and AlphaEvolve illustrate a trajectory of expanding searchable objects while keeping verification cheap. Formal‑math efforts like AlphaProof and Seed‑Prover show similar patterns in theorem‑proving.

The core insight is that verification cost, not candidate quantity, is the bottleneck for true scientific discovery. MLS‑Bench frames the challenge as achieving scalable discovery when verification cannot be scaled.

Conclusion

Current models are efficient engineers that can execute research pipelines, yet they lack the ability to independently propose, validate, and generalize new methods. MLS‑Bench provides a systematic infrastructure for measuring progress toward genuine AI‑driven scientific discovery.

References: [1] When AI builds itself – https://www.anthropic.com/institute/recursive-self-improvement; [2] Nature paper – https://www.nature.com/articles/s41586-025-09833-y; [3] Seed‑Prover – https://arxiv.org/abs/2507.23726; [4] Kimi K3 Tech Blog – https://www.kimi.com/blog/kimi-k3; [5] Qwen3.8‑Max Blog – https://qwen.ai/blog?id=qwen3.8