Search papers, labs, and topics across Lattice.
This paper investigates the effectiveness of constitutional midtraining鈥攊ntegrating principled, values-based content during midtraining鈥攐n the alignment of large language models (LLMs). By employing a 394M-token constitutional corpus and a factorial design, the authors demonstrate that models subjected to constitutional midtraining show significant improvements in alignment generalization and durability, particularly in mitigating blackmail tendencies, compared to a control group. Notably, while the alignment benefits persist through benign fine-tuning, they diminish in scenarios requiring active resistance to pressure or conflict, highlighting the nuanced impact of midtraining interventions on model behavior.
Constitutional midtraining can significantly enhance alignment durability in LLMs, effectively reducing blackmail propensity even after fine-tuning.
Post-training alignment is often shallow, eroding under fine-tuning. Whether midtraining interventions, cleanly isolated from post-training, can produce durable alignment remains untested. We test this via constitutional midtraining: inserting principled, values-based content into midtraining against a replay-only control at 120B scale. Our 394M-token constitutional corpus, built from Anthropic's Constitution, uses a 2x2 factorial design (curriculum ordering x deliberative reasoning) to produce four constitutionally midtrained conditions plus a control, evaluated on self-generated and established benchmarks including alignment under pressure, value conflict resolution, blackmail, and emergent misalignment across three stages: post-midtraining, post-SFT, and post-benign fine-tuning. Constitutionally midtrained models outperform the control on alignment generalization and durability, notably on blackmail: SFT instills a blackmail propensity in all models, but constitutional midtraining blunts it, with the advantage surviving benign fine-tuning (-17.5pp). This durability does not extend to settings requiring active resistance to in-context pressure or conflict, where the advantage attenuates after SFT. The presence of constitutional content at midtraining also matters more than its structure, and constitutional midtraining incurs no cost, on average, on the capabilities we test (MMLU, ARC-Easy, piqa, GSM8K) at any stage. A modest amount of constitutional content at midtraining could therefore yield broad, persistent alignment gains, offering a cheap, complementary addition to SFT-centered pipelines. Code, data, and models are available.
