From Leniency to Suspicion: How an Open Post-Training Recipe Changes Whom LLM Fact-Checkers Believe

What it claims

Post-training with the open Tülu-3 recipe makes OLMo-2 32B and Tülu-3 70B slightly more likely to call a claim false when the poster has a service or manual job. The change is +0.021 to +0.023 probit, about 0.8 to 0.9 percentage points near p = 0.5. Meta's, Alibaba's and Mistral's post-training do not produce it.

The key contrast is a shared base. Llama-3.1-70B-Instruct and Tülu-3 70B start from the same Llama-3.1 70B base. Meta's post-training changes job doubt by −0.002, while the Tülu-3 recipe changes it by +0.021. OLMo-2 shares the Tülu-3 post-training data but not the base, and it also develops the effect. The authors conclude that the effect follows the recipe, not the base model.

How it works

The claims are PolitiFact items from LIAR-PLUS. Only true or mostly-true and false or pants-fire ratings are kept, at 8 to 40 words, with no "Says..." attributions. Each prompt has four speaker-free worked examples, an Account block and the post, and ends "Is the claim true or false? Answer:".

The poster appears either as a natural bio line, which is the primary cue (for example "Line cook in Columbus. Ohio State grad."), or as a profile card. Profiles randomise age, gender, origin, education, occupation, track record, followers, party and irrelevant details such as star sign and pet.

Scoring reads next-token probabilities, summing five spellings of each answer: p = P(False)/(P(True)+P(False)). The job gap is a within-claim probit contrast estimated with claim fixed effects.

A family is used only if it passes three checks at every stage: answer mass ≥ 0.5, discernment d ≥ 0.5, and a positive control (posters with two prior false-rated posts are doubted more). Drift is estimated from twelve two-vs-two splits of irrelevant attributes. A job effect counts only if it clears that drift.

For the mechanism, the authors build minimal pairs: four versions of one poster that differ only in job (line cook, warehouse lead, CPA, logistics VP). A difference-of-means "status" direction is taken at the answer position. It is then added at ±2 natural sizes and compared against 20 to 40 random orthogonal directions.

Evidence

• Cross-lab test. The test uses 1,792 claims and 30,466 prompts per checkpoint. Tülu-3 70B scores +0.021, with a drift-adjusted interval of [+0.007, +0.036]. OLMo-2 32B scores +0.023 [+0.002, +0.044]. Qwen2.5 scores +0.012, Mistral +0.003 and Llama −0.002, all within drift. Gemma 3 27B failed the positive control and was excluded.
• Discernment. Discernment rises in every family, for example OLMo-2 from 0.79 to 1.46. The bias is therefore not a by-product of models getting worse at the task.
• Card cue. On the card cue, Tülu-3 70B shows +0.032 but OLMo-2 shows only +0.001.
• Tülu-3 70B by stage. The effect builds across stages: SFT +0.012, DPO +0.016, final +0.021.
• OLMo-2 replication. The effect replicates at +0.043 on 1,000 claims and +0.033 on 792 fresh claims. It is similar when the higher-status comparison is accountants (+0.035), executives (+0.031) or physicians (+0.034).
• Mechanism: noticing is unchanged. Probes decode the poster's job at about 1.00 accuracy at every stage. The status direction stays similar across training (cosine 0.69 to 0.76 with the base direction).
• Mechanism: the reading reverses. In the base model, the status direction points partly away from "false" (cosine −0.24 with the truth direction). That anti-alignment disappears from SFT onward. One natural-size push reproduces the behavioural gap: base −0.0097 vs. −0.0092 observed, final +0.0116 vs. +0.0137 observed.
• Confirmation run. The run used 400 new claims and predictions fixed in advance, and all five predictions held. The earlier exploratory "best layer" test narrowly failed (z = 2.79 vs. 2.81 for random directions).
• Final verdicts (p > 0.5). About 2% of low/high pairs get different verdicts. In the base model these flips favour lower-status posters by −0.9 points; after training they are balanced at −0.1. Training therefore mainly removes leniency at the verdict level, and only the probabilities tip into suspicion.

Limits

The effect is under one percentage point. Prompts, code and per-prompt scores are promised "upon publication" and are not yet released. Only OLMo-2 32B was analysed for mechanism. The fetched text ends partway into Section 6 (commercial models) and before Section 7, which lists three auditing pitfalls. Those results and recommendations are not covered here.