What specific training signal would help LLMs learn non-local constraint satisfaction instead of just pattern-matching derivations?

Your question sits outside the archive's scope—these preprints address *whether* LLMs can learn physics reasoning and *how to evaluate* that learning, but not how to design training signals for constraint satisfaction in the abstract sense you're asking.

That said, the closest relevant thread is in [AG-2026.04-565], which identifies the specific failure mode: LLMs perform well on "explicit derivations within stable conceptual frames" but fail systematically when tasks require "reorganization of representations under global consistency constraints" [AG-2026.04-565]. This suggests the problem is real and measurable.

Two papers hint at solutions in different ways:

On synthetic data generation: [AG-2026.04-892] developed a pipeline generating over 2,500 synthetic physics problems for fine-tuning, and conducted both supervised and reinforcement learning experiments, analyzing how "reasoning errors evolve during RL and SFT" [AG-2026.04-892]. The RL track is suggestive—reward signals tied to constraint satisfaction rather than token matching might help—but the paper doesn't detail what those rewards were.

On architectural robustness: [AG-2026.01-735] proposes that robust reasoning emerges from non-Abelian gauge symmetry in the network architecture itself, demonstrated on variable-binding tasks where models must maintain consistency across large state spaces [AG-2026.01-735]. The implication is that the *inductive bias* of the architecture matters as much as the training signal.

To directly answer your question about specific training signals for non-local constraints, you'd need work that explicitly compares reward functions or loss terms designed to penalize constraint violations—which these preprints don't provide.