Training Large Language Models (LLMs) to reason often relies on Reinforcement Learning (RL) with task-specific verifiers. However, many real-world reasoning-intensive tasks lack verifiers, despite offering abundant expert demonstrations that remain under-utilized for reasoning-focused training. We introduce RARO (Relativistic Adversarial Reasoning Optimization) that learns strong reasoning capabilities from only expert demonstrations via Inverse Reinforcement Learning. Our method sets up an adversarial game between a policy and a relativistic critic: the policy learns to mimic expert answers, while the critic aims to identify the experts among (expert, policy) answer pairs. Both the policy and the critic are trained jointly and continuously via RL, and we identify the key stabilization techniques required for robust learning. Empirically, RARO significantly outperforms strong verifier-free baselines on all of our evaluation tasks -- Countdown, DeepMath, and Poetry Writing -- and enjoys the same robust scaling trends as RL with verifiers. These results demonstrate that our method effectively elicits strong reasoning performance from expert demonstrations alone, enabling robust reasoning learning even when task-specific verifiers are unavailable. Recent advances in Large Language Models (LLMs) have been driven substantially by improvements in their reasoning abilities. Reasoning enables LLMs to perform deliberate intermediate computations before producing answers to the user queries, proposing candidate solutions and self-corrections. Much of this progress has been enabled via Reinforcement Learning (RL) on verifiable tasks such as mathematics and competitive programming (DeepSeek-AI et al., 2025; Y ang et al., 2025a; Shao et al., 2024; Luo et al., 2025). Notably, recent work has demonstrated that RL with V erifiable Rewards (RL VR) can enable LLMs to develop robust reasoning capabilities without any additional supervision (DeepSeek-AI et al., 2025). A growing body of work further improves the efficiency and stability of such RL algorithms on verifiable tasks, such as DAPO (Y u et al., 2025) and GSPO (Zheng et al., 2025). However, comparatively little attention has been paid to developing reasoning abilities on non-verifiable tasks, where task-specific verifiers are unavailable. Y et, in many impactful and challenging tasks -- such as analytical writing, open-ended research, or financial analysis -- LLM outputs are not directly verifiable due to hard-to-specify criteria, wide variation among acceptable answers, and other practical constraints. A popular approach in these settings is Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022; Rafailov et al., 2023), but they require collecting human preferences beyond demonstration data, which is often a time-consuming and expensive process.

Reasoning models leveraging long chains of thought employ various cognitive skills, such as verification of their answers, backtracking, retrying by an alternate method, and more. Previous work has shown that when a base language model exhibits these skills, training that model further with reinforcement learning (RL) can learn to leverage them. How can we get models to leverage skills that aren't exhibited by base models? Our work, SkillFactory, is a method for fine-tuning models to roughly learn these skills during a supervised fine-tuning (SFT) stage prior to RL. Our approach does not rely on distillation from a stronger model, but instead uses samples from the model itself, rearranged to provide training data in the format of those skills. These "silver" SFT traces may be imperfect, but are nevertheless effective for priming a model to acquire skills during RL. Our evaluation shows that (1) starting from SkillFactory SFT initialization helps a model to generalize to harder variants of a task post-RL, despite lower performance pre-RL; (2) cognitive skills are indeed used by the model; (3) RLed SkillFactory models are more robust to regression on out-of-domain tasks than RLed base models. Our work suggests that inductive biases learned prior to RL help models learn robust cognitive skill use.

Should LLM reasoning live in a separate module, or within a single model's forward pass and representational space? We study dual-architecture latent reasoning, where a fluent Base exchanges latent messages with a Coprocessor, and test two hypotheses aimed at improving latent communication over Liu et al. (2024): (H1) increase channel capacity; (H2) learn communication via joint finetuning. Under matched latent-token budgets on GPT-2 and Qwen-3, H2 is consistently strongest while H1 yields modest gains. A unified soft-embedding baseline, a single model with the same forward pass and shared representations, using the same latent-token budget, nearly matches H2 and surpasses H1, suggesting current dual designs mostly add compute rather than qualitatively improving reasoning. Across GSM8K, ProsQA, and a Countdown stress test with increasing branching factor, scaling the latent-token budget beyond small values fails to improve robustness. Latent analyses show overlapping subspaces with limited specialization, consistent with weak reasoning gains. We conclude dual-model latent reasoning remains promising in principle, but likely requires objectives and training schedules that explicitly shape latent spaces for algorithmic planning.

Masked Diffusion Models (MDMs) as language models generate by iteratively unmasking tokens, yet their performance crucially depends on the inference time order of unmasking. Prevailing heuristics, such as confidence based sampling, are myopic: they optimize locally, fail to leverage extra test-time compute, and let early decoding mistakes cascade. We propose Lookahead Unmasking (LookUM), which addresses these concerns by reformulating sampling as path selection over all possible unmasking orders without the need for an external reward model. Our framework couples (i) a path generator that proposes paths by sampling from pools of unmasking sets with (ii) a verifier that computes the uncertainty of the proposed paths and performs importance sampling to subsequently select the final paths. Empirically, erroneous unmasking measurably inflates sequence level uncertainty, and our method exploits this to avoid error-prone trajectories. We validate our framework across six benchmarks, such as mathematics, planning, and coding, and demonstrate consistent performance improvements. LookUM requires only two to three paths to achieve peak performance, demonstrating remarkably efficient path selection. The consistent improvements on both LLaDA and post-trained LLaDA 1.5 are particularly striking: base LLaDA with LookUM rivals the performance of RL-tuned LLaDA 1.5, while LookUM further enhances LLaDA 1.5 itself showing that uncertainty based verification provides orthogonal benefits to reinforcement learning and underscoring the versatility of our framework. Code will be publicly released.

Leveraging inference-time search in large language models has proven effective in further enhancing a trained model's capability to solve complex mathematical and reasoning problems. However, this approach significantly increases computational costs and inference time, as the model must generate and evaluate multiple candidate solutions to identify a viable reasoning path. To address this, we propose an effective approach that integrates search capabilities directly into the model by fine-tuning it on unpaired successful (learning) and failed reasoning paths (forgetting) derived from diverse search methods. A key challenge we identify is that naive fine-tuning can degrade the model's search capability; we show this can be mitigated with a smaller learning rate. Extensive experiments on the challenging Game-of-24 and Countdown arithmetic puzzles show that, replacing CoT-generated data with search-generated data for offline fine-tuning improves success rates by around 23% over inference-time search baselines, while reducing inference time by 180$\times$. On top of this, our learning and forgetting objective consistently outperforms both supervised fine-tuning and preference-based methods.

Post-training has demonstrated its importance in enhancing the reasoning capabilities of large language models (LLMs). The primary post-training methods can be categorized into supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). SFT is efficient and well-suited for small language models, but it may lead to overfitting and limit the reasoning abilities of larger models. In contrast, RFT generally yields better generalization but depends heavily on the strength of the base model. To address the limitations of SFT and RFT, we propose Unified Fine-Tuning (UFT), a novel post-training paradigm that unifies SFT and RFT into a single, integrated process. UFT enables the model to effectively explore solutions while incorporating informative supervision signals, bridging the gap between memorizing and thinking underlying existing methods. Notably, UFT outperforms both SFT and RFT in general, regardless of model sizes. Furthermore, we theoretically prove that UFT breaks RFT's inherent exponential sample complexity bottleneck, showing for the first time that unified training can exponentially accelerate convergence on long-horizon reasoning tasks.

Diffusion large language models (dLLMs) are emerging as an efficient alternative to autoregressive models due to their ability to decode multiple tokens in parallel. However, aligning dLLMs with human preferences or task-specific rewards via reinforcement learning (RL) is challenging because their intractable log-likelihood precludes the direct application of standard policy gradient methods. While prior work uses surrogates like the evidence lower bound (ELBO), these one-sided approximations can introduce significant policy gradient bias. To address this, we propose the Sandwiched Policy Gradient (SPG) that leverages both an upper and a lower bound of the true log-likelihood. Experiments show that SPG significantly outperforms baselines based on ELBO or one-step estimation. Specifically, SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K, 2.6% in MATH500, 18.4% in Countdown and 27.0% in Sudoku.

Diffusion large language models (dLLMs) generate text through iterative denoising, yet current decoding strategies discard rich intermediate predictions in favor of the final output. Our work here reveals a critical phenomenon, temporal oscillation, where correct answers often emerge in the middle process, but are overwritten in later denoising steps. To address this issue, we introduce two complementary methods that exploit temporal consistency: 1) Temporal Self-Consistency Voting, a training-free, test-time decoding strategy that aggregates predictions across denoising steps to select the most consistent output; and 2) a post-training method termed Temporal Consistency Reinforcement, which uses Temporal Semantic Entropy (TSE), a measure of semantic stability across intermediate predictions, as a reward signal to encourage stable generations. Empirical results across multiple benchmarks demonstrate the effectiveness of our approach. Using the negative TSE reward alone, we observe a remarkable average improvement of 24.7% on the Countdown dataset over an existing dLLM. Combined with the accuracy reward, we achieve absolute gains of 2.0% on GSM8K, 4.3% on MATH500, 6.6% on SVAMP, and 25.3% on Countdown, respectively. Our findings underscore the untapped potential of temporal dynamics in dLLMs and offer two simple yet effective tools to harness them.

Recent advancements in large language models (LLMs) 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 intro 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.

Recent breakthroughs in large language models (LLMs) have effectively improved their reasoning abilities, particularly on mathematical and logical problems that have verifiable answers, through techniques such as supervised finetuning (SFT) and reinforcement learning (RL). Prior research indicates that RL effectively internalizes search strategies, enabling long chain-of-thought (CoT) reasoning, with backtracking emerging naturally as a learned capability. However, the precise benefits of backtracking, specifically, how significantly it contributes to reasoning improvements and the optimal extent of its use, remain poorly understood. In this work, we systematically investigate the dynamics between SFT and RL on eight reasoning tasks: Countdown, Sudoku, Arc 1D, Geometry, Color Cube Rotation, List Functions, Zebra Puzzles, and Self Reference. Our findings highlight that short CoT sequences used in SFT as a warm-up do have moderate contribution to RL training, compared with cold-start RL; however such contribution diminishes when tasks become increasingly difficult. Motivated by this observation, we construct synthetic datasets varying systematically in the number of backtracking steps and conduct controlled experiments to isolate the influence of either the correctness (content) or the structure (i.e., backtrack frequency). We find that (1) longer CoT with backtracks generally induce better and more stable RL training, (2) more challenging problems with larger search space tend to need higher numbers of backtracks during the SFT stage. Additionally, we demonstrate through experiments on distilled data that RL training is largely unaffected by the correctness of long CoT sequences, suggesting that RL prioritizes structural patterns over content correctness. Collectively, our results offer practical insights into designing optimal training strategies to effectively scale reasoning in LLMs.