Recent research has increasingly focused on reconciling the reasoning capabilities of System 2 with the efficiency of System 1. While existing training-based and prompt-based approaches face significant challenges in terms of efficiency and stability, model merging emerges as a promising strategy to integrate the diverse capabilities of different Large Language Models (LLMs) into a unified model. However, conventional model merging methods often assume uniform importance across layers, overlooking the functional heterogeneity inherent in neural components. To address this limitation, we propose Activation-Guided Consensus Merging (ACM), a plug-and-play merging framework that determines layer-specific merging coefficients based on mutual information between activations of pre-trained and fine-tuned models. ACM effectively preserves task-specific capabilities without requiring gradient computations or additional training. Extensive experiments on Long-to-Short (L2S) and general merging tasks demonstrate that ACM consistently outperforms all baseline methods. For instance, in the case of Qwen-7B models, TIES-Merging equipped with ACM achieves a 55.3% reduction in response length while simultaneously improving reasoning accuracy by 1.3 points.

Recent large language models (LLMs) have demonstrated strong reasoning capabilities that benefits from online reinforcement learning (RL). These capabilities have primarily been demonstrated within the left-to-right autoregressive (AR) generation paradigm. In contrast, non-autoregressive paradigms based on diffusion generate text in a coarse-to-fine manner. Although recent diffusion-based large language models (dLLMs) have achieved competitive language modeling performance compared to their AR counterparts, it remains unclear if dLLMs can also leverage recent advances in LLM reasoning. To this end, we propose d1, a framework to adapt pre-trained masked dLLMs into reasoning models via a combination of supervised finetuning (SFT) and RL. Specifically, we develop and extend techniques to improve reasoning in pretrained dLLMs: (a) we utilize a masked SFT technique to distill knowledge and instill self-improvement behavior directly from existing datasets, and (b) we introduce a novel critic-free, policygradient based RL algorithm called diffu-GRPO, the first integration of policy gradient methods to masked dLLMs. Through empirical studies, we investigate the performance of different post-training recipes on multiple mathematical and planning benchmarks. We find that d1 yields the best performance and significantly improves performance of a state-of-the-art dLLM. Our code is released at https://dllm-reasoning.github.io/.

Training agents to operate under strict constraints during deployment, such as limited resource budgets or stringent safety requirements, presents significant challenges, especially when these constraints render the task complex. In this work, we propose a curriculum learning strategy that gradually tightens constraints during training, enabling the agent to incrementally master the deployment requirements. Inspired by self-paced learning techniques in unconstrained reinforcement learning (RL), our approach facilitates a smoother transition to challenging environments by initially training on simplified versions of the constraints and progressively introducing the full deployment conditions. We provide a theoretical analysis using an RL agent in a binary-tree Markov Decision Process (MDP) to demonstrate that our curriculum strategy can accelerate training relative to a baseline approach that imposes the trajectory constraints from the outset.

Improving the reasoning capabilities of large language models (LLMs) typically requires supervised fine-tuning with labeled data or computationally expensive sampling. We introduce Unsupervised Prefix Fine-Tuning (UPFT), which leverages the observation of Prefix Self-Consistency - the shared initial reasoning steps across diverse solution trajectories - to enhance LLM reasoning efficiency. By training exclusively on the initial prefix substrings (as few as 8 tokens), UPFT removes the need for labeled data or exhaustive sampling. Experiments on reasoning benchmarks show that UPFT matches the performance of supervised methods such as Rejection Sampling Fine-Tuning, while reducing training time by 75% and sampling cost by 99%. Further analysis reveals that errors tend to appear in later stages of the reasoning process and that prefix-based training preserves the model's structural knowledge. This work demonstrates how minimal unsupervised fine-tuning can unlock substantial reasoning gains in LLMs, offering a scalable and resource-efficient alternative to conventional approaches.

The proof is an induction on k. Consider the general case p2k+1. It is easy to see that g (x) = ex p2k(x) and g (x) = ex p2k 1(x). By the induction hypothesis, g 0 and therefore g is convex. Thus, the minimum of g is given by its stationary points. It is easy to observe that x = 0 is indeed a stationary point. Thus, minx R g(x) = g(0) = 0, which finishes the proof.

In this section, we give a formal functional definition of the decision-making policies introduced in Section 3. During each task, the agent sequentially observes samples xi [ 1,1] representing realizations of stochastic observations of the current innovation value. A map τ: [ 1,1]N N is a duration (of a decision task) if for all x [ 1,1]N, its value d= τ(x) Nat xdepends only on the first dcomponents x1,x2,...,xd of x = (x1,x2,...); mathematically speaking, if X is a discrete stochastic process (i.e., a random sequence), then τ(X) is a stopping time with respect to the filtration generated by X. This definition reflects the fact that the components x1,x2,... of the sequence x = (x1,x2,...) are generated sequentially, and the decision to stop testing an innovation depends only on what occurred so far. A concrete example of a duration function is the one, mentioned in the introduction and formalized in (4), that keeps drawing samples until the empirical average of the observed values xi surpasses/falls below a certain threshold, or a maximum number of samples have been drawn.

We consider the case where each message fails with probability 1 p and each agent i uses the messages it receives from its neighbors with probability pi.This is equivalent to each agent ireceiving messages from its neighbors with probability pip.Let 1{(i,j) 2 Et}be the indicator random variable that takes value 1 if agent i receives reward value and arm id from agent j at time t and 0 otherwise. We start by proving some useful lemmas. Lemma 1. (Restatement of results from [3]) Let k = Thus we have P Ai(t+1) = k,Nik(t) > k P bµi1(t) µ1 Ci1(t) +P bµik(t) µk +Cik(t) This concludes the proof of Lemma 1. Lemma 2. Let (G) is the clique covering number of graph G. Let k = Let C be a non overlapping clique covering of G. Then we have that k |C| < Nik( ik,C) k. From regret results it follows that regret for this case is greater than the regret for the case where ik,C < k,C for some (or all) i. 13 We analyse the expected number of times agents pull suboptimal arm k as follows, X P bµi1(t) µ1 Ci1(t) +P bµik(t) µk +Cik(t), (29) where (a) follows from the fact that clique covering is non overlapping. This concludes the proof of Lemma 2. Lemma 3. Let di(G) be the degree of agent i in graph G.