Large Language Models (LLMs) are increasingly controlled through system prompts that specify roles, styles, formats, and safety requirements. However, models follow these prompts only implicitly through in-context learning, which can be insufficient for complex or compositional prompts. Existing approaches often require model tuning or response-level reranking, limiting their practicality for lightweight inference-time control. We introduce SyRuP, a decoding-time framework for improving system-prompt adherence while keeping the base LM frozen. SyRuP trains a cross-attention reward head from system-prompt-conditioned preference pairs, treating the system prompt as a separate memory to produce token-level adherence scores. At inference, SyRuP reranks the base LM's top-k candidates by combining base logits with the learned reward signal and an optional contrastive signal capturing system-induced logit shifts. Experiments on system-prompt following benchmarks show that SyRuP consistently outperforms prompting and decoding-time baselines with moderate inference overhead. These results suggest that explicit token-level guidance is an effective and practical mechanism for reliable system-prompt following.
Discrete diffusion models offer a simple and stable likelihood-based framework for sequence generation, recently extended to any-length settings via token insertion. Principled reward-guided fine-tuning for any-length discrete diffusion, however, remains largely unexplored. We introduce Fine-Tuning Any-Length Discrete Diffusion for Adaptive Decoding (A2D2), a unified framework for reward-guided fine-tuning of any-length discrete diffusion models via joint optimization of the insertion and unmasking policies together with a quality-based inference schedule. We derive the Radon-Nikodym derivative for the joint insertion-unmasking path measures, enabling theoretically guaranteed convergence to the intractable reward-tilted sequence distribution without requiring target samples. Building on this, we establish unmasking and insertion quality as tractable approaches for minimizing decoding error and introduce the Adaptive Joint Decoding (AJD) loss, which provably yields the optimal path measure that generates the reward-tilted distribution. Empirically, A2D2 improves reward optimization while enhancing generation flexibility and accuracy over prior fixed-length fine-tuning and inference-time guidance methods.
Vision-Language-Action (VLA) policies remain brittle in long-horizon and high-uncertainty control, where one-pass action decoding provides limited inference-time deliberation. Explicit chain-of-thought can increase reasoning depth, but introduces token latency and an indirect text-to-action interface. We propose MPCoT, a reward-guided multi-path latent reasoning framework that initializes M hypotheses, refines them for K weight-tied steps, and softly aggregates them before action decoding. A training-only path-preference objective combines expert-trajectory consistency, frozen Qwen3-VL progress scoring, and endpoint-success feedback to align the latent path scorer with downstream execution quality. MPCoT preserves the original 8-step action interface, generates zero reasoning tokens, and exposes configurable inference controls (K, M). Under matched protocols on LIBERO and CALVIN, MPCoT improves long-horizon performance, with ablations confirming depth-width effects, confidence-weighted aggregation, and reward-guided path supervision.