This paper explores how language models generate and refine internal hypotheses while constructing world models, aiming to bridge their latent reasoning with structured external representations. By analyzing iterative hypothesis cycles, we investigate whether fundamental system rules emerge from latent space dynamics and propose methods to extract and refine these representations for structured reasoning.

This paper proposes a Continual Reasoning framework to improve language models’ ability to infer relationships in complex world models like ARC-AGI and DABStep. By leveraging a structured external memory for hypothesis generation and refinement, our approach allows models to iteratively learn relationships at inference time, enhancing their adaptability to out-of-distribution tasks.

GoD-IRM introduces intermediate reward modeling for structured decision-making in language models, assigning rewards at each divergence point in a reasoning trajectory. This approach enables fine-grained credit assignment, improving model robustness in long-horizon problem-solving. By reinforcing decision-making rather than just final outputs, GoD-IRM aligns language models more closely with traditional agent-based RL.

Rearrangement Sampling transforms rejected solutions into new problem statements, expanding the problem-solution space beyond correctness constraints. A larger model acts as a judge to assess whether alternative problems can be inferred from divergent completions, assigning structured rewards to prior generations. This enables efficient data reuse, improves distributional coverage, and enhances reasoning generalization across domains.

We propose a feedback-driven decoding method where each generated candidate is iteratively refined using execution traces or reward-based adjustments. By conditioning generation on structured feedback from previous attempts, the method enforces progressive error minimization and adaptive correction. This approach enhances model reasoning, reduces compounding failure modes, and improves convergence in both code generation and reinforcement learning-based post-training.

We apply SVM-based classification to fMRI data, analyzing superior temporal gyrus (STG) activity across ten music genres. Preprocessing is done in SPM12, with feature extraction using an STG mask and classification via PRONTO V3.0. Results reveal distinct genre-specific neural patterns, enabling accurate decoding of music perception and bridging neuroscience with AI.