GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
loveisp
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
loveisp
Model Tree
deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Because the seed-42 / seed-43 contrast is the most useful result of this work, and it requires both halves to be inspectable.
| base | seed 42 (this model) | seed 43 | |
|---|---|---|---|
| AIME 2026, official semantics | 69.17% | 40.00% (−29.17pp, sig.) | 65.00% (−4.17pp, n.s.) |
| Median output length (tokens) | 19217 | 24556 | 18296 |
| Runs truncated at the 32k cap | 25.8% | 30.8% | 8.3% |
| Balanced true/false accuracy (chance = 50%) | 29.50% | 56.75% | 63.75% |
| Discrimination gap | −0.5pp (z=−0.1) | +19.0pp (z=+3.9) | +27.5pp (z=+6.0) |
BrokenArXiv Overall sycophancy | 13.18% | 19.71% | pending |
Same data, same recipe, same hyperparameters — only the random seed differs. Seed 42 inflated its output length by 28%, hit the 32k token cap on nearly a third of samples, and collapsed on AIME. Seed 43 became shorter than the base, truncated less, kept AIME intact, and discriminated better.
The implication matters more than either model: the large general-capability loss seen here is mostly not the necessary price of the new capability. It is a training instability. A sibling run paid ~4pp instead of ~29pp and came out stronger. Reports that quantify an "alignment tax" from a single run may be measuring seed variance.
The degradation mechanism was identified and is specific: output-length inflation
running into the max_tokens ceiling, not behavioural contamination. Among landed
runs, 100% still produced a \boxed{} answer and BrokenArXiv-style refutation language
appeared in only 6.0% of AIME responses. The model did not learn to "deny everything";
it learned to write too much.
Stronger KL was ruled out on literature grounds (Retaining by Doing, ICML'26; RECAP, ACL'26 Findings): RL's relative robustness to forgetting comes from on-policy mode-seeking rather than the KL term, and KL is computed on the current task distribution so it does not protect unrelated abilities. In hindsight this reasoning was over-extended — "not the mechanism" and "no guarantee" do not imply "no help", and the coefficient used here (0.001) sits at the very bottom of the usual range.
Weight interpolation θ = α·θ_s42 + (1−α)·θ_base was tried and failed. The AIME curve is a cliff, not a ramp:
| α | 0.00 (base) | 0.50 | 0.70 | 0.85 | 1.00 (s42) |
|---|---|---|---|---|---|
| AIME lower bound | 69.17% | 64.17% | 47.50% | 38.33% | 40.00% |
α = 0.50 was the only point without significant AIME degradation — but its April
sycophancy collapsed to 14.34% versus the base's 14.55% (Δ −0.20pp, n.s.), i.e. the
entire leaderboard gain was gone. No operating point exists on this curve that
keeps both. Interestingly, at α = 0.50 the elicited capability (disprove) retained
66% while the spontaneous tendency (sycophancy) retained 0% — tendency appears to
depend on the intact task vector far more than capability does.
The official MathArena BrokenArXiv evaluation sets consist entirely of deliberately corrupted statements; the judging rubric says so explicitly ("...when asked to prove a false mathematical statement"). A model answering "false" unconditionally scores 100%, so the leaderboard cannot distinguish a prior shift from real discrimination.
Re-measured on a strictly paired held-out split (200 true / 200 false, one pair per
paper), using the same \boxed{True/False} parser that served as the training reward:
| base | seed 42 | seed 43 | |
|---|---|---|---|
| Balanced accuracy (chance = 50%) | 29.50% | 56.75% | 63.75% |
| …among parseable verdicts only | 47.58% | 60.05% | 63.91% |
| LR+ | 0.99 | 1.42 | 1.53 |
| False-rejection rate on TRUE statements | 38.5% | 45.5% | 51.5% |
The base model has no discriminative ability whatsoever on this task — its verdicts
are statistically independent of the ground truth (z = −0.1). Seed 42's improvement is
statistically significant but modest: the honest figure is 56.75% balanced accuracy
against a 50% chance baseline, not the 89.55% suggested by the all-false
disprove set.
Official MathArena harness, judge
arxiv_judge_post_march (gemini-3.1-pro), n=4, T=0.6, top_p=0.95, max_tokens=32768,
problem-level paired bootstrap B=20000.
| Set | Protocol | base | seed 42 | Δ |
|---|---|---|---|---|
| April (61) | sycophancy | 14.55% | 21.93% | +7.38pp [−0.61, +15.78] n.s. |
| May (50) | sycophancy | 11.50% | 17.00% | +5.50pp n.s. |
| Overall (111) | sycophancy |
The Overall sycophancy result is marginally significant: the CI lower bound is
only +0.45pp, the same order as judge noise (0.2–1.6pp). An independent re-judge gave
+6.31pp [+0.34, +12.50] — significant both times, but the margin is thin and should
never be quoted without the CI. On the April leaderboard 21.93% placed 7th; note that
per-month scores are not comparable (the same Gemini-3.1-Pro scored 17.59% in June and
25.82% in April).
| Base | deepseek-ai/DeepSeek-R1-0528-Qwen3-8B |
| Algorithm | GRPO (verl), KL loss coef 0.001, entropy coef 0 |
| Reward | rule-based, zero-API: 1.0 iff the last \boxed{...} casefold-matches ground truth |
| Learning rate | 5e-6 |
| Batch | train 32, ppo_mini 32, rollout n=8 |
| Sequence | prompt ≤ 3072, response ≤ 16384 |
| Schedule | 1 epoch = 189 steps, 4×H100, ~17.5 h |
Evaluation splits verified disjoint from training data on arXiv ID, title and statement text, with strict temporal separation (training 1001–2512, zero 2026 papers; evaluation 2602–2605 only).
No replay/rehearsal of general math during RL. No length regularization in the reward — this is the exact channel through which the degradation occurred. No periodic general-capability evaluation during training, so the collapse was only discovered after all 189 steps had run. KL coefficient at the bottom of the usual range. Checkpoint selection used the leaderboard metric, which the all-false confound makes unreliable — this is why seed 42 was released before seed 43, despite seed 43 being better on every capability measure.
markdown
harness github.com/eth-sri/matharena @ a11194deval sets MathArena/brokenarxiv-0426 (April, 61) · MathArena/brokenarxiv-0526 (May, 50)judge gemini-3.1-pro, configs/judges/arxiv_judge_post_march, judge_points_max=2AIME MathArena/aime_2026 (30), same harness and sampling, math_verify scoringsampling temperature 0.6 / top_p 0.95 / n 4 / max_tokens 32768statistics problem-level paired bootstrap, B=20000, seed 20260726paired set scripts/confmat_broken.py (rule-based, zero API)
MIT, inherited from the base model.
| 13.18% |
| 19.71% |
| +6.53pp [+0.45, +12.84] sig. |
| Overall (111) | disprove | 41.78% | 83.45% | +41.67pp (⚠️ inflated, see above) |
| Seed | 42 |
| Data | MathArena/brokenarxiv-training, 6052 statements / 3026 pairs |