To Backtrack or Not to Backtrack: When Sequential Search Limits Model Reasoning

Recent advancements in large language models have significantly improved their reasoning abilities, particularly through techniques involving search and backtracking. Backtracking naturally scales test-time compute by enabling sequential, linearized exploration via long chain-of-thought (CoT) generation. However, this is not the only strategy for scaling test-time compute: parallel sampling with best-of-n selection provides an alternative that generates diverse solutions simultaneously. Despite the growing adoption of sequential search, its advantages over parallel sampling--especially under a fixed compute budget remain poorly understood. In this paper, we systematically compare these two approaches on two challenging reasoning tasks: CountDown and Sudoku. Surprisingly, we find that sequential search underperforms parallel sampling on CountDown but outperforms it on Sudoku, suggesting that backtracking is not universally beneficial. We identify two factors that can cause backtracking to degrade performance: (1) training on fixed search traces can lock models into suboptimal strategies, and (2) explicit CoT supervision can discourage "implicit" (non-verbalized) reasoning. Extending our analysis to reinforcement learning (RL), we show that models with backtracking capabilities benefit significantly from RL fine-tuning, while models without backtracking see limited, mixed gains. Together, these findings challenge the assumption that backtracking universally enhances LLM reasoning, instead revealing a complex interaction between task structure, training data, model scale, and learning paradigm.

$Π$-NeSy: A Possibilistic Neuro-Symbolic Approach

In this article, we introduce a neuro-symbolic approach that combines a low-level perception task performed by a neural network with a high-level reasoning task performed by a possibilistic rule-based system. The goal is to be able to derive for each input instance the degree of possibility that it belongs to a target (meta-)concept. This (meta-)concept is connected to intermediate concepts by a possibilistic rule-based system. The probability of each intermediate concept for the input instance is inferred using a neural network. The connection between the low-level perception task and the high-level reasoning task lies in the transformation of neural network outputs modeled by probability distributions (through softmax activation) into possibility distributions. The use of intermediate concepts is valuable for the explanation purpose: using the rule-based system, the classification of an input instance as an element of the (meta-)concept can be justified by the fact that intermediate concepts have been recognized. From the technical side, our contribution consists of the design of efficient methods for defining the matrix relation and the equation system associated with a possibilistic rule-based system. The corresponding matrix and equation are key data structures used to perform inferences from a possibilistic rule-based system and to learn the values of the rule parameters in such a system according to a training data sample. Furthermore, leveraging recent results on the handling of inconsistent systems of fuzzy relational equations, an approach for learning rule parameters according to multiple training data samples is presented. Experiments carried out on the MNIST addition problems and the MNIST Sudoku puzzles problems highlight the effectiveness of our approach compared with state-of-the-art neuro-symbolic ones.

Verification Learning: Make Unsupervised Neuro-Symbolic System Feasible

The current Neuro-Symbolic (NeSy) Learning paradigm suffers from an over-reliance on labeled data. If we completely disregard labels, it leads to less symbol information, a larger solution space, and more shortcuts-issues that current Nesy systems cannot resolve. This paper introduces a novel learning paradigm, Verification Learning (VL), which addresses this challenge by transforming the label-based reasoning process in Nesy into a label-free verification process. VL achieves excellent learning results solely by relying on unlabeled data and a function that verifies whether the current predictions conform to the rules. We formalize this problem as a Constraint Optimization Problem (COP) and propose a Dynamic combinatorial Sorting (DCS) algorithm that accelerates the solution by reducing verification attempts, effectively lowering computational costs to the level of a Constraint Satisfaction Problem (CSP). To further enhance performance, we introduce a prior alignment method to address potential shortcuts. Our theoretical analysis points out which tasks in Nesy systems can be completed without labels and explains why rules can replace infinite labels, such as in addition, for some tasks, while for others, like Sudoku, the rules have no effect. We validate the proposed framework through several fully unsupervised tasks including addition, sort, match, and chess, each showing significant performance and efficiency improvements.

Noise to the Rescue: Escaping Local Minima in Neurosymbolic Local Search

Deep learning has achieved remarkable success across various domains, largely thanks to the efficiency of backpropagation (BP). However, BP's reliance on differentiability poses challenges in neurosymbolic learning, where discrete computation is combined with neural models. We show that applying BP to Godel logic, which represents conjunction and disjunction as min and max, is equivalent to a local search algorithm for SAT solving, enabling the optimisation of discrete Boolean formulas without sacrificing differentiability. However, deterministic local search algorithms get stuck in local optima. Therefore, we propose the Godel Trick, which adds noise to the model's logits to escape local optima. We evaluate the Godel Trick on SATLIB, and demonstrate its ability to solve a broad range of SAT problems. Additionally, we apply it to neurosymbolic models and achieve state-of-the-art performance on Visual Sudoku, all while avoiding expensive probabilistic reasoning. These results highlight the Godel Trick's potential as a robust, scalable approach for integrating symbolic reasoning with neural architectures.

Hyperspherical Energy Transformer with Recurrent Depth

Transformer-based foundation models have achieved unprecedented success with a gigantic amount of parameters and computational resources. Yet, the core building blocks of these models, the Transformer layers, and how they are arranged and configured are primarily engineered from the bottom up and driven by heuristics. For advancing next-generation architectures, it demands exploring a prototypical model that is amenable to high interpretability and of practical competence. To this end, we take a step from the top-down view and design neural networks from an energy minimization perspective. Specifically, to promote isotropic token distribution on the sphere, we formulate a modified Hopfield energy function on the subspace-embedded hypersphere, based on which Transformer layers with symmetric structures are designed as the iterative optimization for the energy function. By integrating layers with the same parameters, we propose \textit{Hyper-Spherical Energy Transformer} (Hyper-SET), an alternative to the vanilla Transformer with recurrent depth. This design inherently provides greater interpretability and allows for scaling to deeper layers without a significant increase in the number of parameters. We also empirically demonstrate that Hyper-SET achieves comparable or even superior performance on both synthetic and real-world tasks, such as solving Sudoku and masked image modeling, while utilizing fewer parameters.

Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions

In recent years, masked diffusion models (MDMs) have emerged as a promising alternative approach for generative modeling over discrete domains. Compared to autoregressive models (ARMs), MDMs trade off complexity at training time with flexibility at inference time. At training time, they must learn to solve an exponentially large number of infilling problems, but at inference time, they can decode tokens in essentially arbitrary order. In this work, we closely examine these two competing effects. On the training front, we theoretically and empirically demonstrate that MDMs indeed train on computationally intractable subproblems compared to their autoregressive counterparts. On the inference front, we show that a suitable strategy for adaptively choosing the token decoding order significantly enhances the capabilities of MDMs, allowing them to sidestep hard subproblems. On logic puzzles like Sudoku, we show that adaptive inference can boost solving accuracy in pretrained MDMs from $<7$% to $\approx 90$%, even outperforming ARMs with $7\times$ as many parameters and that were explicitly trained via teacher forcing to learn the right order of decoding.

T-SCEND: Test-time Scalable MCTS-enhanced Diffusion Model

We introduce Test-time Scalable MCTS-enhanced Diffusion Model (T-SCEND), a novel framework that significantly improves diffusion model's reasoning capabilities with better energy-based training and scaling up test-time computation. We first show that na\"ively scaling up inference budget for diffusion models yields marginal gain. To address this, the training of T-SCEND consists of a novel linear-regression negative contrastive learning objective to improve the performance-energy consistency of the energy landscape, and a KL regularization to reduce adversarial sampling. During inference, T-SCEND integrates the denoising process with a novel hybrid Monte Carlo Tree Search (hMCTS), which sequentially performs best-of-N random search and MCTS as denoising proceeds. On challenging reasoning tasks of Maze and Sudoku, we demonstrate the effectiveness of T-SCEND's training objective and scalable inference method. In particular, trained with Maze sizes of up to $6\times6$, our T-SCEND solves $88\%$ of Maze problems with much larger sizes of $15\times15$, while standard diffusion completely fails.Code to reproduce the experiments can be found at https://github.com/AI4Science-WestlakeU/t_scend.

Evaluating SAT and SMT Solvers on Large-Scale Sudoku Puzzles

Modern SMT solvers have revolutionized the approach to constraint satisfaction problems by integrating advanced theory reasoning and encoding techniques. In this work, we evaluate the performance of modern SMT solvers in Z3, CVC5 and DPLL(T) against a standard SAT solver in DPLL. By benchmarking these solvers on novel, diverse 25x25 Sudoku puzzles of various difficulty levels created by our improved Sudoku generator, we examine the impact of advanced theory reasoning and encoding techniques. Our findings demonstrate that modern SMT solvers significantly outperform classical SAT solvers. This work highlights the evolution of logical solvers and exemplifies the utility of SMT solvers in addressing large-scale constraint satisfaction problems.

Surveying the space of descriptions of a composite system with machine learning

Multivariate information theory provides a general and principled framework for understanding how the components of a complex system are connected. Existing analyses are coarse in nature -- built up from characterizations of discrete subsystems -- and can be computationally prohibitive. In this work, we propose to study the continuous space of possible descriptions of a composite system as a window into its organizational structure. A description consists of specific information conveyed about each of the components, and the space of possible descriptions is equivalent to the space of lossy compression schemes of the components. We introduce a machine learning framework to optimize descriptions that extremize key information theoretic quantities used to characterize organization, such as total correlation and O-information. Through case studies on spin systems, Sudoku boards, and letter sequences from natural language, we identify extremal descriptions that reveal how system-wide variation emerges from individual components. By integrating machine learning into a fine-grained information theoretic analysis of composite random variables, our framework opens a new avenues for probing the structure of real-world complex systems.

Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning

Autoregressive language models, despite their impressive capabilities, struggle with complex reasoning and long-term planning tasks. We introduce discrete diffusion models as a novel solution to these challenges. Through the lens of subgoal imbalance, we demonstrate how diffusion models effectively learn difficult subgoals that elude autoregressive approaches. We propose Multi-granularity Diffusion Modeling (MDM), which prioritizes subgoals based on difficulty during learning. On complex tasks like Countdown, Sudoku, and Boolean Satisfiability Problems, MDM significantly outperforms autoregressive models without using search techniques. For instance, MDM achieves 91.5\% and 100\% accuracy on Countdown and Sudoku, respectively, compared to 45.8\% and 20.7\% for autoregressive models. Our work highlights the potential of diffusion-based approaches in advancing AI capabilities for sophisticated language understanding and problem-solving tasks.