GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
idealab-cs2
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
idealab-cs2
Model Tree
Qwen/Qwen3-4B-Thinking-2507
Input Modalities
Output Modalities
Supported Functionality
mean(z) − 0.5·std(z), a
disagreement guard (Coste et al. 2023) that down-weights answers the two scorers disagree on.Reappraisals are judged pairwise by Llama-3.1-70B-Instruct (both A/B orderings), calibrated against the human ratings (93% agreement on clearly separated pairs; larger judges agreed less and showed position bias). These numbers are on the study's 6 vignettes, which the model was also trained on, so read them as evidence about reward-model quality on a narrow task, not a general capability ranking. Intervals are a vignette-cluster bootstrap (the contests cluster on 6 scenarios, so plain iid intervals are optimistic).
On the 6 vignettes, head-to-head, both orders:
| opponent | win-rate | n | CI-low |
|---|---|---|---|
| GPT-4-0314 | 0.87 | 720 | 0.78 |
| DeepSeek-R1-671B | 0.67 | 240 | 0.55 |
CI-low is above the 0.5 tie line in both cases: a 4B single-pass model out-reappraises both GPT-4-0314 and R1-671B pairwise on these scenarios.
Off-distribution it is a different story. On fresh scenarios the model never trained on, it does not beat R1-671B and sits at roughly parity with a strong open 72B (Qwen2.5-72B, ~0.49). Off-distribution reappraisal quality tracks base-model capacity more than our training, so the in-distribution wins do not generalize to "a 4B beats the frontier."
Why the committee reward helps: the gain is from combining two different-mechanism scorers (a discriminative RM and a generative one), not from adding a second model. An RM-heavy 0.7/0.3 mix scores worse (0.76); equal weighting is what helps, because the two scorers make different mistakes and rewarding only their agreement penalizes answers that game either one alone.
The model was trained with this system prompt:
markdown
You are helping someone reduce a negative emotion they feel in a short interpersonal scenario byoffering an alternative interpretation of the situation (a 'reappraisal' / 'rethinking'). Directyour response at the person in the scenario in second person. Limit your response to two sentencesmaximum. Do not list emotions; output only the reappraisal text.
The user turn is
markdown
SCENARIO:\n{scenario}\n\nWrite a reappraisal of this scenario in two sentences maximum, addressing the person in second person.
<think>...</think> followed by the reappraisal; only the text after
</think> is the answer.
The reward carries a format gate penalizing outputs that are empty or longer than two sentences (so the policy cannot inflate reward through verbosity). Training was healthy: reward rising, KL bounded (~0.60), entropy alive, no reward-std collapse.
The 6 vignettes' GPT-4-0314 reappraisals and the frontier-opponent answers are held out and used only for evaluation; training uses the vignette prompts (the task) and the human-rating-based reward models. Both committee scorers are Qwen-family, distinct from the Llama-3.1-70B evaluation judge, and the generative model's references are real human reappraisals, never the held-out GPT-4 answers or the judge's outputs. Any pooled or synthetic training text is n-gram-checked against the held-out eval answers before use.
Research artifact for the study of RLHF, reward modeling, and computational emotion regulation. It writes short cognitive reappraisals of negative interpersonal situations. It is not a clinical or mental-health tool and must not be used as one. Quality is validated only on the narrow reappraisal task and the evaluation above.