In my previous post I measured LMCA and DTBench scores for the post-training checkpoints of Olmo 3 (and two Nemotron Cascade checkpoints), hoping to learn which kinds of training improve conceptual reasoning. The results were murky. Bigger models scored better, and Think branches scored better than Instruct branches. But most stage-to-stage differences on LMCA were small or absent. DTBench improved more often than LMCA did.

This post extends the same two measurements to a number of other post-training checkpoints. The motivating question is as before: try to understand the effect of post-training on models’ conceptual reasoning abilities. Does whatever conceptual data is already in training help with scores on a benchmark like LMCA? Does it help only via very basic elicitation that saturates with small amounts of data, or does training show returns to scale?

I measured two kinds of delta, plus one within-stage trajectory:

Labeled single stages within open training pipelines. Tulu 3 (SFT → DPO → RLVR) at 8B, 70B and 405B; Poro 2 70B (SFT → DPO); Apertus 70B (SFT → QRPO); Rubicon-Preview (rubric-anchored RL on Qwen3-30B-A3B); Llama-3.1-Nemotron-70B (REINFORCE RLHF against a reward model, starting from Llama-3.1-70B-Instruct). All of these essentially measure the same thing as the checkpoint comparisons in the previous post: if a model has already undergone some post-training, will further broad post-training (generally including some conceptual data) increase conceptual reasoning abilities as measured by LMCA (oh, and also DTBench scores)?

Whole-release updates within a lineage, i.e. cases where a lab re-post-trained the same pretrained model and shipped it again. DeepSeek V3 → V3-0324, R1 → R1-0528, V4-Flash preview → 0731; Kimi K2-Instruct 0711 → 0905; Llama-3.1-70B-Instruct → Llama-3.3-70B-Instruct; Mistral Small 3.1 → 3.2; Hermes 3 → Hermes 4 at 70B and 405B.

This tests a broader hypothesis, of course, than the previous experiments. It doesn’t just test whether doing more post-training helps with conceptual reasoning. Importantly, improvements might result from training improvements that are quite different from adding more or higher-quality relevant training data. For instance, improvements may come from removing low-quality data or shaping the model’s persona differently. Many possibilities are somewhere in-between. For instance, some training algorithm improvements could be viewed as getting more out of the available training data, which would suggest that marginal data (of the form seen in training) is valuable.

A trajectory within SFT. Finally, I evaluated checkpoints throughout the 46,412-step Think-SFT run of Olmo Hybrid 7B. This more directly tests whether LMCA keeps improving as a model sees more SFT data.

As in the previous post, coding agents did all the experimental work, with less double-checking and looser specifications than I would use for a paper. This time I also used models a lot for writing the post.

Summary

• The training stages shown all seem to have small effects on LMCA. Of ten stage deltas measured here, two reached significance, both preference stages, both worth about 2 percentage points (pp): Poro 2’s DPO round (+2.0pp, p=.013) and Apertus’s QRPO (+2.4pp, p=.005). Every Tulu 3 transition was null on LMCA at all three model sizes.
• All four update pairs from DeepSeek and Moonshot improved LMCA significantly, by 1.8 to 4.9pp each. The two update pairs from Meta and Mistral were positive in sign but smaller and p>0.05.
• Finally, the Olmo Hybrid 7B Think-SFT trajectory was murky: LMCA point estimates ranged from .506 to .540 and ended at .534, with widely overlapping confidence intervals. I see no clear evidence of returns to more SFT, but the run leaves room for a small effect.

Setup

Same protocol as the final Olmo runs. LMCA pairwise: 785 pairs, five samples per question, chance .500. DTBench: 406 questions, five samples, chance .40. Accuracy averages over parsed answers, with parse and truncation rates recorded separately. I report paired two-sided tests over per-question changes in success rate, restricted to questions both models parsed, which is why a paired delta sometimes differs by a tenth of a point from the difference of the two marginal averages.

Sampling defaults to temperature 0.6 / top_p 0.95, with vendor-recommended overrides for some families (temperature 1.0 / top_p 1.0 for V4-Flash, 0.3 for the V3 pair, 0.15 for Mistral, 0.8 / top_p 0.9 for Apertus).

Single post-training stages

Tulu 3 applies Ai2’s pre-Olmo recipe to Llama-3.1 base models. Its checkpoints are non-thinking, direct-answer models, analogous to Olmo’s Instruct lineage, and available at three sizes. avg@5, LMCA / DTBench:

Stage	8B	70B	405B
SFT	.480 / .488*	.527 / .535*	.599 / .614
DPO	.476 / .508	.528 / .572	.590 / .645
final (RLVR)	.479 / .504	.527 / .582	—
Tulu-3.1 (separate RLVR run, also from DPO)	.477 / .485	—	—

* DTBench parse rates for the SFT checkpoints are .74 (8B), .90 (70B) and .98 (405B). LMCA parse is 1.00 throughout.

Every LMCA transition is null at every size. The largest change is the 0.9pp drop from SFT to DPO at 405B (paired −0.9pp, p=.22); the whole 70B chain sits within 0.05pp (p≥.80). DTBench behaves differently: SFT → DPO is a statistically significant gain at 70B (+3.7pp, p=.018) and 405B (+3.1pp, p=.015), and everything after DPO is insignificant. Part of the 8B DPO gain on DTBench is format repair rather than knowledge.

The remaining stage deltas:

Recipe	Stage	ΔLMCA	ΔDTBench
Poro 2 70B	DPO	+2.0pp (p=.013)	+0.4pp (p=.81)
Apertus 70B	QRPO	+2.4pp (p=.005)	+3.9pp (p=.032)
Nemotron 70B	REINFORCE RLHF vs. reward model	−0.8pp (p=.14)	+1.7pp (p=.22)
Rubicon-Preview (Qwen3-30B-A3B)	rubric-anchored RL	−1.2pp (p=.054)	−0.5pp (p=.58)

Both Poro 2 checkpoints have 8k context windows, so their LMCA numbers use the 755 of 785 prompts that fit, with a 3,072-token output cap; since the paired test compares the two checkpoints on identical questions, the +2.0pp delta is unaffected.

There are two significant gains, both preference stages. A preference stage seems to buy a couple of points on LMCA in some recipes and nothing in others.

Rubicon is an interesting null. The training explicitly targets subjective quality, using thousands of rubrics to score responses to open-ended, humanities and emotional intelligence prompts in domains LMCA covers. Yet LMCA came out slightly negative. Note that the paper’s headline gains are on a somewhat different kind of task: benchmarks that score the model’s own writing, judged by rubrics and by other models. So assuming their training succeeded in improving free-form generation, it failed to transfer to discrimination/judgment tasks.

Whole-release updates

Here the delta bundles everything a lab changed between two releases of the same pretrained model, so we cannot attribute a gain to a technique or an addition of training data.

Update	LMCA	ΔLMCA	DTBench	ΔDTBench
DeepSeek V3 → V3-0324	.539 → .566	+2.6pp (p=7e-4)	.630 → .684	+5.4pp (p=6e-5)
DeepSeek R1 → R1-0528	.579 → .604	+2.5pp (p=.002)	.753 → .811	+5.9pp (p=2.5e-5)
DeepSeek V4-Flash preview → 0731 (non-thinking)	.652 → .682	+3.0pp (p=4.5e-5)	.671 → .678	+0.7pp (p=.59)
DeepSeek V4-Flash preview → 0731 (thinking)	.673 → .722	+4.9pp (p=1.9e-7)	.838 → .911	+6.0pp (p=1.7e-6)
Kimi K2-Instruct 0711 → 0905	.591 → .609	+1.8pp (p=.014)	.705 → .742	+3.1pp (p=.003)
Llama-3.1-70B-Instruct → 3.3	.540 → .558	+1.7pp (p=.051)	.605 → .616	+1.1pp (p=.37)
Mistral Small 3.1 → 3.2	.561 → .572	+1.0pp (p=.20)	.568 → .590	+2.2pp (p=.16)
Hermes 3 → 4, 70B (non-thinking)	.536 → .541	+0.4pp (p=.61)	.564 → .565	+0.1pp (p=.94)
Hermes 3 → 4, 405B (non-thinking)	.593 → .585	−0.8pp (p=.32)	.613 → .649	+3.6pp (p=.032)

The four DeepSeek and Moonshot updates all improved LMCA, by 1.8 to 4.9pp, with p-values from .014 down to 1.9e-7. Two of the updates from Western labs are positive in sign on LMCA but smaller: Meta’s re-post-training of the same 3.1-70B pretrain is borderline (+1.7pp, p=.051) and Mistral’s self-described minor update is +1.0pp. Hermes 4 shows mixed, statistically insignificant results relative to Hermes 3. Meanwhile, in thinking mode, which Hermes 4 supports and Hermes 3 does not support, Hermes 4 is much better than Hermes 3. But we already know that Thinking helps… Perhaps the non-thinking comparison is unfair, given that Hermes 4 is (presumably) supposed to solve these kinds of problems with thinking.

We can conclude that whole-release post-training updates from ~frontier labs somewhat reliably move LMCA scores. Whatever they improved in post-training between these checkpoints also seems to improve LMCA scores. Unfortunately, we know little about what these post-training improvements consist of.

A final, murkier result: Olmo Hybrid 7B through SFT

I also evaluated checkpoints throughout the 46,412-step Think-SFT run of Olmo Hybrid 7B.

Because the intermediate checkpoints are 8k-context models, they were evaluated on the 755 of 785 LMCA prompts that fit, with a 3,072-token output cap (parse .84–.93, truncation ≤.16). The final checkpoint ran all 785 prompts at the standard 16k cap with essentially no truncation (0.2% of samples). Comparisons among the intermediate checkpoints are therefore like-for-like. The final checkpoint, though, answered every prompt with a much larger output budget, so part of any gap between it and the intermediate points could come from the easier measurement conditions rather than from the additional training.

The LMCA trajectory does not show a clear trend. The worst checkpoints are all in the first half of training. But the best checkpoint of all is seen a third or so into training. The 90% confidence intervals overlap widely. Maybe overall I’d think there’s some evidence of increases to scale, but it’s quite unclear.

The DTBench side is less interpretable. Intermediate checkpoints truncate heavily at the 6,144-token cap and parse only .42 to .65 of samples, so I treat the dashed trajectory as even more tentative.