Pleias CTO Pierre-Carl Langlais joins an active X thread to connect GLM-5.2's agentic strengths to heavy reinforcement learning on synthetic environments plus recursive loops that generate and critique their own tasks, with the whole idea sparked by one oddly shaped training diagram.
Based on the bouba shape, my guess would be hard synth/rl env scaling with recursive generative design+eval.
Has anyone done any speculation on the training recipe of GLM 5.2? Beyond extensive RL, we know it's (at least?) a new midtrain ("GLM-5.2 is trained with IndexShare from mid-training with 128K sequence length") with arch changes.
Langlais and other engineers are treating the bouba-shaped visual as a hint toward scaled rollout environments and self-improving task synthesis, an angle the released GLM-5.2 docs do not confirm or deny.
If the speculated recipe holds, the model's already strong Terminal-Bench and SWE-bench numbers could translate into longer autonomous runs on real repos without constant human steering.
Users welcome GLM-5 training methods using RL scaling and synthetic tasks because they let the open ecosystem generate data and avoid billion-dollar collection costs.
No Digg Deeper questions have been answered for this story yet.
Ok answer in plain sight in the original GLM 5.0 paper. They just synthetize RL env at scale.
Has anyone done any speculation on the training recipe of GLM 5.2? Beyond extensive RL, we know it's (at least?) a new midtrain ("GLM-5.2 is trained with IndexShare from mid-training with 128K sequence length") with arch changes.
@Dorialexander I feel like these generated tasks are likely subpar.. how much manual review by human is done I wonder
Ok answer in plain sight in the original GLM 5.0 paper. They just synthetize RL env at scale.
i have gone from kvetching about mass generated synth tasks being subpar to "wait, rejection sampling + careful judge heuristic filtering from a giant set of mid envs can be strategically kino actually?"
@Dorialexander I feel like these generated tasks are likely subpar.. how much manual review by human is done I wonder
synth envs can give you a broad lens on the types of things a model is good or bad at, and clever filters + diagnostics can highlight gaps you wouldn't normally find i.e sorting tasks by what GLM saturates but GPTmini flails at lets you ask, whats the correlated meta problem?
i have gone from kvetching about mass generated synth tasks being subpar to "wait, rejection sampling + careful judge heuristic filtering from a giant set of mid envs can be strategically kino actually?"
Should settle part of my research program for this summer.
Ok answer in plain sight in the original GLM 5.0 paper. They just synthetize RL env at scale.
actually, along with other closed labs signals, doesn't seem great news for custom rl env sellers.
here it looks a lot like "GPT mini isnt nearly as robust to irrelevant information being in context"! you can of course stratify the use of heuristic filters in other ways. e.g, LM as judge + correlated judge/verifier disagreement can highlight envs wrong at *the semantic level*
synth envs can give you a broad lens on the types of things a model is good or bad at, and clever filters + diagnostics can highlight gaps you wouldn't normally find i.e sorting tasks by what GLM saturates but GPTmini flails at lets you ask, whats the correlated meta problem?
@yacineMTB They’re likely seeded + critical part is redirecting the annotation at the meta-level.
@Dorialexander I feel like these generated tasks are likely subpar.. how much manual review by human is done I wonder
@yacineMTB (likely through the OPD recipe, since they mention it, "efficiently merging more than ten expert models into the final model")
@Dorialexander when you say RL env scaling, you mean total volume of RL envs right?
Based on the bouba shape, my guess would be hard synth/rl env scaling with recursive generative design+eval.
@yacineMTB yeah and diversity/combinations.
the judge (or judges) don't even have to be "consistently right" per se; the value here can come from isolating *where judgements systematically skew at all* vs verifier scores across a large n, which can predict how dubious a verifier's rules are relative to the task semantics
here it looks a lot like "GPT mini isnt nearly as robust to irrelevant information being in context"! you can of course stratify the use of heuristic filters in other ways. e.g, LM as judge + correlated judge/verifier disagreement can highlight envs wrong at *the semantic level*
and, conversely, much better news for the open ecosystem that can maybe shortcut a billion-dollars data building capability by generating it all. though you'll still need hard skills.
@yacineMTB @Dorialexander You’d think a Chinese company could scale up manual review faster using south Asian resources
@Dorialexander what i dont understand wh they are using deepseek v3.2 achitecture+ some glm stuff instead of the newer version. cuz of the deeoseek v4 architecure compexity?
@yacineMTB @Dorialexander you can probably just do hq phase at the end i'd imagine
@JayooHwang Skill issue all along.
@kalomaze Might be a skill issue but I self-hypnotise about how good my filtering is after spending too much time on it
@Dorialexander So all you need is a bunch of cracked Tsinghua researchers designing agentic workflows for creating RL envs?
@Kinch_ahoy @yacineMTB Not in a short time frame. Anthropic env build up took 1-2 years.