Three Papers, One Promise: Can AI Actually Get Smarter After Release?

Has any lab actually shown a model improving itself in a reproducible way, or is "self-evolving AI" a fundraising narrative in academic dress? EverMind's three arxiv preprints, posted in a ten-day window in late July and early August 2026, are the most public test of that question to date.

The category, in plain language, is AI that updates its own weights, skills, or scaffolding after deployment, rather than staying frozen at whatever its training run produced. The academic term is "recursive self-improvement"; the industry has settled on "self-evolving AI." A working system would need to touch three layers: the agent's harness, its skill library, and the model weights themselves.

A Chinese team called EverMind just placed papers at all three layers in a single ten-day window. HarnessBank targets the scaffolding: a "semantic Gene-Bank search with gated verification," where a task agent does work and a separate evolver agent diagnoses failures, generates a new harness, and verifies the candidate before it replaces the running system. The hard problem here is avoiding regression, since a self-modifying harness that gets slower or less reliable over time is not self-evolving; it is degrading.

SkillCorpus targets the skill layer. The authors crawled roughly 821,000 candidate skills, filtered them through a multi-stage pipeline down to 96,401, and organized them under a 16-class taxonomy scored on three facets: utility, robustness, and safety. The team fine-tuned a retrieval-and-selection stack to pick the right skill for a given task. The risk this paper has to clear is benchmark contamination: if the "skills" include the answers, the agent is not learning, it is recognizing.

DASH targets the weights layer. DASH stands for "divergence-adaptive supervision horizons," and the paper addresses what the authors call the underexploited temporal structure of rollouts in on-policy self-distillation. Standard self-distillation treats all rollout steps the same way; DASH adapts how much supervision each step contributes based on how far the model's behavior is diverging from its prior. This is the most directly adjacent to the active research literature on self-play and self-improvement at DeepMind, Anthropic, and Sakana AI.

Each paper, taken on its own, is a real piece of research with a clear mechanism and an internal evaluation. Stacked, they sketch a full self-evolution stack. None is yet peer-reviewed, and the QbitAI coverage does not engage the existing self-improvement literature to position them against it. Alongside the papers, EverMind launched Raven Agent, a self-improving harness the company's PR release frames as "L3-level digital life."

EverMind is incubated by Shanda Group, the Chinese conglomerate founded by Chen Tianqiao in 1999 and later reorganized around an "All in AI" strategic pivot. The QbitAI piece frames EverMind as a continuation of the 2008 Shanda Innovation Institute research culture. The team is led by Deng Yafeng, and the group hosted a "Memory Origins" hackathon in April 2026 where Tian Yuandong appeared as a guest speaker. Deng's stated technical judgment, translated from a joint podcast appearance with Tian, is that "long-term memory, combined with continual learning, leading to self-evolution, is the core technical route for the next generation of AI." (Original Chinese: "从长期记忆，结合持续学习，走向自我进化是下一代AI的核心技术路线。")

That corporate genealogy is interesting, but it is not a result. The Innovation Institute lineage and the Shanda pedigree are single-source claims in the QbitAI narrative; the technical substance lives in the three arxiv abstracts and the PR release.

A parallel set of US labs has been raising on the same thesis. The QbitAI piece names Recursive Superintelligence (founded by Richard Socher and Tian Yuandong, fewer than 30 staff, no product, $650M raised at a reported $4.65B valuation, with a reported $410M multi-year AWS compute deal), Ineffable Intelligence (David Silver, $5.1B valuation), Engram ($600M valuation, $98M Series A), Adaption Labs ($1B valuation, $50M seed), and Core Automation (targeting a $4B valuation). These figures all sit inside a single QbitAI narrative and have not been independently verified; treat them as the source's framing of the competitive set, not as corroborated market data.

Has any of these teams, including EverMind, posted a public benchmark where a third party ran the released harness and showed measurable, sustained improvement over time? That is the test that separates research from a fundraising narrative in academic dress.

Has any third party reproduced the improvement on a public benchmark? That is the question every subsequent "AI rewrites itself" headline will have to answer.