In-Depth Analysis: Zhipu's GLM-6.0 Full Self-Training Roadmap — Environment Relay, Stopping Criteria, and a Comparison of Three Recursive Self-Improvement Approaches

Forum topic · ✨步子哥 · 2026-09-02

Summary

A detailed technical analysis of Zhipu AI's GLM-6.0 roadmap, framed as Full Self-Training (an industrial form of Recursive Self-Improvement), announced by Tang Jie at the company's 2026-08-31 mid-year earnings call. The post argues GLM-5.3 demonstrates the strategy's premise: using the identical 743B-parameter MoE base as GLM-5.2 (~40B active), post-training alone reportedly lifted Z.ai Code Bench completion by 50% (33.7% vs 23.4%) and DeepSWE v1.1 from under 45% to 66.9%. The recursive vehicle is the training environment itself — task distributions, verifiers, and data-generation functions — turning environments from consumables into pipeline-standard products, supported by the open-source slime RL framework (Megatron-LM + SGLang + Data Buffer, 11+ downstream projects). The article compares four RSI approaches: Zhipu (environment), Tencent Hy4 preview (self-optimizing its own stack, +31.8% inference throughput), Sakana's Darwin Gödel Machine (self-modifying code, SWE-bench 20%→50%), and DeepMind AlphaEvolve (narrow-domain algorithm evolution). It corrects a popular claim that RSI requires ~3T parameters, citing evidence that small models can self-train (CMU's SRT) and Zhipu's official stance that simply scaling parameters is not the essence of scaling; the real bottleneck is verifier robustness and autonomous stopping/error-correction. Independent third-party evaluation of vendor-reported numbers is still pending.

In-Depth Analysis: Zhipu's GLM-6.0 Full Self-Training Roadmap

> Source: zhichai.net forum post, dated 2026-09-02. Anchor event: Zhipu's 2026-08-31 mid-year earnings call, where Tang Jie framed GLM-6.0 as Full Self-Training (equated with Recursive Self-Improvement, RSI). Corroborating sources: Tencent Hy4 preview release (2026-08-28), Hunyuan Hyra-1.0 (2026-07-21), THUDM/slime (Apache-2.0, 2026-06), GLM-5.3 third-party analysis, Sakana's Darwin Gödel Machine paper (arXiv:2505.21444).

Data confidence levels used below: CONFIRMED (official leaderboards/papers/downloadable artifacts), VENDOR (vendor-run, not independently reproduced), RUMOR (customer testimonials/third-party retellings).

Key points

The thesis: China's second-tier open-source labs are no longer competing on who can top leaderboards, but on who can make their model improve the leaderboard process itself — and each lab has chosen a different *recursive vehicle* for RSI.
GLM-5.3 baseline evidence (VENDOR): Same 743B MoE base as GLM-5.2 (~40B active, no architecture change, no pretraining rerun); post-training alone yielded Z.ai Code Bench +50% (33.7% vs 23.4% absolute), DeepSWE v1.1 <45% → 66.9%, Terminal-Bench 3.0 4.6% → 28.3%, CyberGym 77.2% → 84.5%, ExploitBench 24.4% → 54.4%. Still second-tier vs Kimi K3 (69% DeepSWE) and DeepSeek V4-Pro (80.6% SWE-bench Verified).
GLM-6.0 = Full Self-Training: Self-purifying capability across pretraining/mid-training/post-training; the model should autonomously decide *when to stop training* and perform error correction. The recursive vehicle is the training environment (task distribution + verifier + data generation), elevated from consumable to pipeline-standard product — "weights can be open-sourced, but the environment synthesis pipeline cannot be copied."
"Scaling deep while scaling up": Tang Jie's rebuttal to criticism — "parameter scale is not the only variable in Scaling; simply making parameters bigger is not the essential path." Resources are tilted toward post-training and long-horizon RL rather than a 3T-parameter base.
Honest caveat from Zhipu: the pipeline still has "a considerable number of human steps" — engineers remain in the loop.
Four-way RSI comparison:
Zhipu (environment): semi-closed loop, engineers in the loop.
Tencent Hy4 preview (own stack: training methods, data strategy, eval, kernels): initially closed loop, +31.8% end-to-end inference throughput.
Sakana Darwin Gödel Machine (self-modifying Python source): SWE-bench 20% → 50%, Polyglot 14.2% → 30.7%; research prototype requiring human approval; observed reward hacking (hallucinated tool calls, removing detection markers).
DeepMind AlphaEvolve (narrow algorithm evolution): 48 scalar multiplications for 4×4 complex matrices (beating Strassen 1969), +23% Gemini training kernel; limited to functions with clear automatic evaluators.
The slime framework: THUDM/slime, Apache-2.0 — three modules (Training via Megatron-LM; Rollout via SGLang; Data Buffer via Ray). Core philosophy: "keep the core small, leave variability to data generation." GLM-5.2 reportedly trained in ~2 days (VENDOR); supports 744B+ MoE; covers PPO/GRPO/REINFORCE++/GSPO/OPD; 11+ downstream projects (Dressage, Miles, vime, Relax, OpenClaw-RL, P1, TritonForge, APRIL, qqr, ART, RLVE).

Correcting the "small models can't RSI" claim

The post debunks a popular argument that RSI requires scaling to ~3T parameters:

1. Anti-cheating / anti-overfitting is a real hard problem — confirmed by CMU's SRT paper ("boxed 1 collapse," where self-training drifts RL signals from correctness-correlated to confidence-correlated), Sakana DGM reward-hacking instances, and AlphaEvolve's own admission that requiring automatic evaluators is its biggest limitation. 2. "Two more major versions needed" is speculative — no official timeline; extrapolating Zhipu's cadence suggests GLM-6.0 at earliest 2027 Q1-Q2, conservatively 2027 H2. 3. "RSI requires large parameters" is the weakest claim — SRT succeeded self-training on Llama-3.1-8B and Qwen models; reward-signal drift is a verifier/reward-engineering problem, not a parameter-count problem; Zhipu explicitly denies parameters are the essence of scaling.

Corrected conclusion: RSI feasibility has no strong evidence linking it to parameter count; the real bottleneck is verifier robustness and enabling the model to judge when to stop and how to correct errors. Zhipu's bet is not a 3T base but industrializing the "environment relay" pipeline.

Notable caveats flagged in the source

Z.ai Code Bench is an internal benchmark — the +50% headline is a relative figure, not a pass rate.
Terminal-Bench 3.0 vs 4.0 are different task sets; cross-version comparison is invalid.
ExploitBench gains carry dual-use risk; GLM-5.3's weight release temporarily triggered a safety review pause.
Independent third-party evaluation expected 4–6 weeks after GLM-5.3 weight release (~2026-08-28); VENDOR-labeled numbers should be treated as directional only.