SQAM Fine-Tunes Flow Policies With a Scalar Adjoint

SQAM fine-tunes flow policies with a scalar adjoint, beating baselines by 18 to 35 points on hard OGBench domains.

Researchers propose Q-learning with Scalar Adjoint Matching (SQAM) to fine-tune flow policies against a learned value function without per-step vector-Jacobian products. A closed-form scalar adjoint scales the final-action value gradient by flow time, and a value penalty constrains the critic at actions the policy generates. The later arXiv report adds that the design is motivated by diagonal concentration of pretrained flow-policy velocity Jacobians. On the four hardest OGBench domains, SQAM's success rate exceeds the strongest baseline by 18 to 35 percentage points. Fine-tuning a vision-language-action policy on a real bimanual robot improved all three tasks over supervised fine-tuning. The two reports agree on those results and do not conflict on any stated figure.

• SQAM (Q-learning with Scalar Adjoint Matching) fine-tunes flow policies against a learned value function.
• A closed-form scalar adjoint scales the final-action value gradient by flow time and avoids per-step vector-Jacobian products.
• The later report motivates the scalar adjoint by diagonal concentration of pretrained flow-policy velocity Jacobians.
• A value penalty is applied on policy-generated actions to keep the critic controlled.
• On the four hardest OGBench domains, success exceeds the strongest baseline by 18 to 35 percentage points.
• Fine-tuning a vision-language-action policy on a real bimanual robot beat supervised fine-tuning on all three tasks.
• Sources are Hugging Face daily papers (2026-10-06) and arXiv cs.AI/cs.LG/cs.CL (2026-10-07); they do not disagree on the reported figures.