We present a discrete diffusion-based generative model for text generation using Glauber dynamics from statistical physics. Our main insight is that instead of trying to train a discrete state space diffusion model using Glauber dynamics with a uniform transition kernel as the forward process, one can set up an “energy function” based on pretrained causal/masked language models, which, when viewed as the stationary distribution, allows us to significantly improve the quality of the generated text. Using UL2 as our pretrained models and modifying and incorporating it into our diffusion pipeline, we obtain significantly better perplexities than prior diffusion-based text generative models and are competitive with the perplexities of GPT-2-medium and GPT-2-large for comparable model sizes. Furthermore, our models outperform prior diffusion models and GPT-2 style auto-regressive models on some zero-shot common sense reasoning tasks as well as some planning/search tasks.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Synthesizing an editable 3D scene from a single RGB image is central to content creation, embodied-agent data generation, and AR/VR, yet remains challenging to achieve both high-fidelity reconstruction and convenient interactive editing. Existing geometry-based pipelines produce high-quality 3D results but are typically hard to refine without rerunning the full process, while LLM-driven procedural systems enable interactive tool use but are mostly text-driven and lack precise metric 3D understanding from images. We present SceneLM, a language-model-based framework that grounds 3D scene synthesis in visual evidence by recovering an executable metric 3D layout directly from a single image. Given an RGB image (and camera intrinsics when available), SceneLM outputs a JSON-form layout specifying each object’s category, 3D center, size, and discretized yaw, and then deterministically executes this layout with a tool suite to instantiate, place, and edit objects for iterative refinement. To train metric layout recovery at scale, we curate five datasets covering diverse indoor, outdoor, and tabletop scenes and convert heterogeneous 3D annotations into a unified instruction-tuning format. To improve numerical stability and metric accuracy while preserving the text interface, we augment autoregressive JSON generation with a lightweight geometry prediction branch and dual supervision. Experiments show that SceneLM substantially improves single-image 3D layout estimation over strong open and proprietary MLLM baselines, and yields higher-quality end-to-end scene generation in geometric consistency, physical plausibility, semantic alignment, and realism.
Large language models (LLMs) often struggle with complex reasoning tasks due to their limitations in addressing the vast reasoning space and inherent ambiguities of natural language. We propose the Mixture-of-Search-Agents (MOSA) paradigm, a novel approach leveraging the collective expertise of multiple LLMs to enhance search-based reasoning. MOSA integrates diverse reasoning pathways by combining independent exploration with iterative refinement among LLMs, mitigating the limitations of single-model approaches. Using Monte Carlo Tree Search (MCTS) as a backbone, MOSA enables multiple agents to propose and aggregate reasoning steps, resulting in improved accuracy. Our comprehensive evaluation across four reasoning benchmarks demonstrates MOSA’s consistent performance improvements over single-agent and other multi-agent baselines, particularly in complex mathematical and commonsense reasoning tasks.
A key challenge for large language models is token cost per query and overall deployment cost. Clinical inputs are long, heterogeneous, and often redundant, while downstream tasks are short and high stakes. We study budgeted context selection, where a subset of document units is chosen under a strict token budget so an off-the-shelf generator can meet fixed cost and latency constraints. We cast this as a knapsack-constrained subset selection problem with two design choices, unitization that defines document segmentation and selection that determines which units are kept.We propose RCD, a monotone submodular objective that balances relevance, coverage, and diversity. We compare sentence, section, window, and cluster-based unitization, and introduce a routing heuristic that adapts to the budget regime. Experiments on MIMIC discharge notes, Cochrane abstracts, and L-Eval show that optimal strategies depend on the evaluation setting. Positional heuristics perform best at low budgets in extractive tasks, while diversity-aware methods such as MMR improve LLM generation. Selector choice matters more than unitization, with cluster-based grouping reducing performance and other schemes behaving similarly. ROUGE saturates for LLM summaries, while BERTScore better reflects quality differences.
Theory of Mind (ToM), the ability to infer others’ mental states from behavior, is pivotal for developing machines with human-level social intelligence. Existing methods endowing LLMs with ToM fall into two paradigms: training-free methods and those repurposing ToM evaluation benchmarks as training data for RL-based fine-tuning. However, training-free methods fail to internalize the augmented ToM into the LLMs. Meanwhile, using evaluation benchmarks as training sources is conceptually problematic and, in practice, results in narrow in-domain overfitting rather than robust ToM. To address the lack of training resources within the ToM community and to empower LLMs with robust ToM, we introduce ToM-Synth, a factorial combinatorial synthesis framework of 6912 social units. This framework enables the systematic synthesis of ToM data, yielding a training dataset of 27,648 instances, termed ToM-Synth-27K. Utilizing ToM-Synth-27K for RL fine-tuning, experimental results demonstrate consistent and significant improvements across models of varying families and scales on ToM, Emotional Intelligence, and Social Commonsense benchmarks. Furthermore, we observe concurrent enhancements in IQ-related tasks (math, science, logic) and effective performance scaling with increasing data scale.
Current code generation evaluation measures functional correctness on well-formed inputs that satisfy all input preconditions. This paradigm has a critical limitation: task descriptions often leave these preconditions implicit, while evaluation filters out inputs that violate them. As a result, generated code may achieve high pass@k scores while failing to enforce the preconditions that the task actually requires. To address this gap, we introduce **ContractEval**, a benchmark for evaluating whether generated code enforces such preconditions—commonly referred to as contracts. Built on HumanEval+ and MBPP+, ContractEval consists of 364 tasks, each with three components: (i) descriptions reconstructed to explicitly state the contracts, (ii) test cases synthesized through a neuro-symbolic pipeline that pairs an LLM with an SMT solver to evaluate whether generated code satisfies these contracts, and (iii) reference code combined with contracts. Using ContractEval to evaluate five representative open-source code LLMs, we reveal a stark disparity between functional correctness and contract satisfaction. Under standard prompting, these models achieve pass@1 of 75-82% with 0% contract satisfaction. Even when contracts are explicitly stated in the prompt, the satisfaction rate reaches only 23-41%. This indicates that current LLMs struggle to satisfy contracts in their generated code, establishing contract satisfaction as a crucial and previously overlooked axis of code generation quality. Our code is available at https://github.com/suhanmen/ContractEval.
Knowledge Graphs (KGs) provide structured and interpretable representations of real-world entities and relations. While dynamic KGs attempt to capture real-time changes, they typically treat updates as independent facts. This overlooks a critical challenge: a factual, localized update can contradict and invalidate previously correct knowledge, requiring revisions beyond the localized update to maintain KG consistency. Many of these inconsistencies arise from events whose effects propagate through relational dependencies, necessitating coordinated multi-hop reasoning rather than isolated changes. To address this, we introduce a model-agnostic framework for cascading KG update identification that leverages conformal prediction to provide reliable uncertainty guarantees over the cascade as a whole, accounting for dependencies among multi-hop update candidates. Building on this foundation, we further develop a graph-based KG update scoring framework that integrates large language models (LLMs) to enrich event representations with world knowledge. Experiments on two newly constructed real-world datasets, designed to reflect scenarios where events necessitate coordinated multi-hop updates, demonstrate that our framework establishes a strong baseline while offering calibrated confidence estimates, providing an effective solution for event-driven KG consistency restoration.
Simul-COMET: A Quality Metric for Simultaneous Interpretation in Distant Language Pair Considering Word Order Difference
PDF ↗In simultaneous interpretation (SI), interpreters perform real-time translation by segmenting the source speech into chunks and translating them in the order they appear.Since surface-matching metrics such as BLEU correlate poorly with human evaluations, translation quality is often evaluated using neural metrics that measure semantic similarity, such as COMET.However, while SI translation ideally exhibits high monotonicity, COMET tends to assign higher scores to offline translations with long-distance reordering, because it is trained on such offline translation data.To address this gap, we propose Simul-COMET, a variation of COMET adapted for SI evaluation specifically designed for monotonicity.We train Simul-COMET on the SI-style translation data, which was converted from the offline translation of the COMET training data by leveraging large language models.In English–Japanese translation experiments, we demonstrate that Simul-COMET assigns higher scores to SI-style translations than to offline ones.Moreover, Simul-COMET shows stronger alignment with evaluation scores provided by professional interpreters than the original COMET.Simul-COMET is available at https://github.com/kosuked/simul-comet.
Traditional reinforcement learning from human feedback (RLHF) optimizes policies on fixed training inputs, limiting the diversity of learning signals. We propose JODP (Joint Optimization of Data and Policy), a framework where the evolving policy model generates improved variants of training problems to enhance its own learning. While training problems remain fixed, JODP optimizes how they are presented: the policy generates specification hints that guide rollout generation, then learns to reproduce the discovered high-reward behaviors without the hints. This "if you can solve it with a hint, learn to solve it without one" principle creates a co-evolutionary dynamic where better policies discover better specifications, which enable further policy improvement. JODP operates as a plug-and-play enhancement to existing algorithms: specifications are selected via UCB bandits for exploration-exploitation balance, used only during training rollouts, and discarded at deployment. Through evaluation on safety alignment tasks, we demonstrate consistent improvements with GRPO, RLOO, and REINFORCE++, allowing 4B models to approach 8B model performance using less than 1% additional computational overhead.
From Fragments to Facts: A Curriculum-Driven DPO Approach for Generating Hindi News Veracity Explanations
PDF ↗In an era of rampant misinformation, generating reliable news explanations is vital, especially for underrepresented languages like Hindi. Lacking robust automated tools, Hindi faces challenges in scaling misinformation detection. To bridge this gap, we propose DeFactoX, a novel framework integrating Direct Preference Optimization (DPO) with Curriculum learning to align machine-generated explanations with human reasoning. Fact-checked explanations from credible sources serve as preferred responses, while LLM outputs highlight system limitations and serve as non-preferred responses. At the core of this framework lies Hin-DPO, an enhanced variant of DPO that enriches the loss function with two novel parameters, Actuality and Finesse, enhancing explanation quality and consistency. Experiments with LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework’s effectiveness in generating coherent, contextually relevant explanations.
Vision-language models (VLMs) trained via reinforcement learning with verifiable reward (RLVR) have shown notable progress in scaling test-time compute effectively. In this work, we investigate how synthesized RL data can further improve RLVR. To this end, we propose SynthRL—a scalable and guaranteed pipeline for automatic data scaling in reasoning-oriented RL training. SynthRL comprises three key stages: (1) selecting seed questions with appropriate distribution, (2) augmenting them into more challenging variants while preserving the original answers, and (3) a guaranteed verification stage that ensures near-perfect correctness and difficulty enhancement. Our empirical experiments demonstrate SynthRL’s scalability and effectiveness. When applied to the MMK12 dataset, SynthRL synthesizes over 3.3K additional verifiable, challenging questions from approximately 8K seed samples. Models trained with our synthesized data achieve consistent gains across five out-of-domain visual math reasoning benchmarks, with a significant improvement over baseline models trained on seed data alone. Notably, detailed analysis reveals that the gains are more pronounced on the most challenging evaluation samples, highlighting SynthRL’s effectiveness in eliciting deeper and more complex reasoning patterns.
Unlocking the Edge deployment and ondevice acceleration of multi-LoRA enabled one-for-all foundational LLM
PDF ↗Deploying large language models (LLMs) on smartphones poses significant engineering challenges due to stringent constraints on memory, latency, and runtime flexibility. In this work, we present a hardware-aware framework for efficient on-device inference of a LLaMA-based multilingual foundation model supporting multiple use cases on Samsung Galaxy S24 and S25 devices with SM8650 and SM8750 Qualcomm chipsets respectively. Our approach integrates application-specific LoRAs as runtime inputs to a single frozen inference graph, enabling dynamic task switching without recompilation or memory overhead. We further introduce a multi-stream decoding mechanism that concurrently generates stylistic variations—such as formal, polite, or jovial responses—within a single forward pass, reducing latency by up to 6×. To accelerate token generation, we apply Dynamic Self-Speculative Decoding (DS2D), a tree-based strategy that predicts future tokens without requiring a draft model, yielding up to 2.3× speedup in decode time. Combined with quantization to INT4 and architecture-level optimizations, our system achieves 4–6× overall improvements in memory and latency while maintaining accuracy across 9 languages and 8 tasks. These results demonstrate practical feasibility of deploying multi-use-case LLMs on edge devices, advancing the commercial viability of Generative AI in mobile platforms.
Multilingual large language models (LLMs) can answer questions in many languages, but how they internally reason across languages remains poorly understood. In this work, we study multilingual reasoning through a decision-making perspective to investigate how multilingual reasoning unfolds in multilingual LLMs using aligned multiple-choice questions from the mMMLU benchmark. By formulating a controlled setup, presenting the same question in different languages, and tracking the model’s decision from the first token to the final answer choice, we can directly compare how reasoning trajectories evolve across languages. We first demonstrate that, at the representation level, different languages share highly similar activation spaces; however, subtle divergences emerge as decisions propagate through the transformer layers. We then model answer selection as a stepwise trajectory, revealing where language-specific signals arise. These patterns are further confirmed by quantifying deviations along these trajectories, highlighting layers where multilingual processing deviates or converges. Our work provides a controlled, layer-resolved view of multilingual reasoning, shedding light on how LLMs balance shared conceptual understanding with language-specific decision-making.
The "alignment tax" of post-training is typically framed as a drop in task accuracy. We show it also involves a severe loss of calibration, making models overconfident, less reliable, and model outputs less diverse. We demonstrate that this trade-off can be navigated effectively via a simple post-hoc intervention: interpolating between a model’s weights before and after alignment. Crucially, this is not a strict trade-off. We find that the process consistently reveals Pareto-optimal interpolations—models that improve accuracy beyond both parents while substantially recovering the calibration lost during alignment. Our work demonstrates that simple model merging provides a computationally efficient method for mitigating the full scope of the alignment tax, yielding models that are more capable and more reliable.
Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling
PDF ↗Reasoning is a core capability of large language models (LLMs), yet how multi-step reasoning is learned and executed remains unclear. We study this question in a controlled cellular-automata (1dCA) framework that excludes memorization by using disjoint training and test rules. Given a short state sequence, the model is required to infer the hidden local rule and then chain it to predict multiple future steps. Our evaluation shows that LLMs largely fail to reliably solve a natural-language proxy of the proposed task. We find that most neural architectures trained from scratch can learn rule inference and achieve high next-step accuracy, but performance drops sharply as the required number of intermediate reasoning steps increases. Experiments show that increasing model depth is crucial, and extending effective depth via recurrence, memory, or test-time compute improves results but remains bounded. Code is available on github: https://github.com/RodkinIvan/associative-recurrent-memory-transformer/tree/ACT.
Language model (LM) evaluators that generate chain-of-thought (CoT) reasoning are widely used for the assessment of LM responses. Simultaneously, increasing LMs’ "thinking" time through scaling test-time compute has proven to be an effective technique for solving challenging problems in domains such as math and code. This raises a natural question: can an LM’s evaluation capability also be improved by scaling test-time compute? To answer this, we investigate employing reasoning models - LMs that natively generate long CoT reasoning - as evaluators. We explore scaling evaluation-time compute by using reasoning models to evaluate both the overall candidate response (i.e., outcome evaluation) and the individual reasoning steps within it (i.e., process evaluation). We observe that evaluator performance improves monotonically with the number of reasoning tokens generated, mirroring trends seen in LM reasoning. Furthermore, we use these more accurate evaluators to rerank multiple generations, and demonstrate that spending more compute at evaluation time can be as effective as increasing compute during generation for improving an LM’s problem-solving performance.
PolyAudio: Advancing Multi-Audio Reasoning in Large Audio Language Models with Interleaved Multi-Audio Contexts
PDF ↗Large Audio Language Models have shown impressive performance on single-clip audio language tasks such as automatic speech recognition, captioning, and sound event recognition. Yet, their ability to reason over interleaved multi-audio contexts-where answering a query requires relating information across multiple audio clips-remains limited. We present PolyAudio, a LALM built on Audio Flamingo 3 that targets multi-audio understanding via instruction tuning rather than massive-scale pre-training, and PolyAudio-Instruct, a high-quality instruction-tuning dataset consisting of 1.3M+ QA pairs, spanning over 14 task subsets to empower multi-audio understanding and reasoning. PolyAudio uses an explicit interleaved representation with clip indexing to encourage faithful grounding and reduce ambiguity in multi-clip references. We evaluate PolyAudio on a diverse suite of multi-audio benchmarks alongside standard single-audio tasks. PolyAudio achieves strong performance on multi-audio reasoning, outperforming competitive baselines that are also often limited to reasoning over up-to 2 audio clips, while preserving robust single-clip performance. Overall, our results suggest that precise, academic-scale multi-audio instruction tuning can unlock advanced cross-clip reasoning capabilities, enabling more capable audio-centric assistants.
Relation extraction is a fundamental task in information extraction. Still, existing supervised approaches rely heavily on large-scale annotated data, limiting their applicability in domain-specific and low-resource scenarios. Prompt-based methods with large language models provide a parameter-efficient alternative; however, their performance is susceptible to prompt design, which often requires extensive domain expertise and heuristic trial-and-error. We propose REPO, a reinforcement learning-based automated prompt optimization framework for domain relation extraction. REPO formulates prompt construction as a structured, sequential decision-making problem, optimizing prompt quality through interaction with a black-box LLM. To enable efficient and stable optimization, we introduce a two-stage framework comprising an initial prompt-construction stage that generates semantically grounded candidates and a DRL-based refinement stage that iteratively improves prompts within a constrained, domain-aware action space. We further design a composite evaluation metric that integrates extraction accuracy and semantic consistency to serve as a dense reward signal. Extensive experiments on multiple relation extraction datasets across medical, financial, legal, and news domains demonstrate that REPO consistently outperforms existing prompt-based methods and supervised baselines. Ablation studies further confirm the effectiveness and robustness of the proposed DRL-based prompt optimization strategy. Our code is available at https://github.com/dddong2-star/REPO.
Hallucination detection remains a significant challenge for large language models. Existing agentic applications rely on LLMs to self-assess the factuality of their outputs using single-step “LLM-as-a-judge” prompts. However, even when equipped with ground truth information, current LLMs still fall short in detecting hallucinations, and this one-shot evaluation offers neither the transparency nor the granularity needed to diagnose where and why the detection fails. To address this gap, we introduce PROBE (Process-based Benchmark for Hallucination Detection), a comprehensive benchmark that breaks down hallucination detection into four critical steps: claim decomposition, evidence finding, evidence evaluation, and hallucination localization, and evaluates each step individually. PROBE consists of 12,000 test cases across three task types—summarization, question answering, and style transfer. Critically, we demonstrate that when hallucination detection is treated as a multi-step process, all models achieve considerably better performance. Through extensive evaluation, we show that current LLMs struggle chiefly with evidence finding, and that finetuning on our released training data substantially improves performance on this step. PROBE represents a significant step toward more transparent, diagnosable, and robust hallucination detection systems.
Recent advancements in large language models (LLMs) have catalyzed the rise of reasoningintensive inference paradigms, where models perform explicit step-by-step reasoning before generating final answers. While such approaches improve answer quality and interpretability, they incur substantial computational overhead due to the prolonged generation sequences. In this paper, we propose Tandem, a novel collaborative framework that synergizes large and small language models (LLMs and SLMs) to achieve high-quality reasoning with significantly reduced computational cost. Specifically, the LLM serves as a strategic coordinator, efficiently generating a compact set of critical reasoning insights. These insights are then used to guide a smaller, more efficient SLM in executing the full reasoning process and delivering the final response. To balance efficiency and reliability, Tandem introduces a cost-aware termination mechanism that adaptively determines when sufficient reasoning guidance has been accumulated, enabling early stopping of the LLM’s generation. Experiments on mathematical reasoning and code generation benchmarks demonstrate that Tandem reduces computational costs by approximately 40% compared to standalone LLM reasoning, while achieving superior or competitive performance. Furthermore, the sufficiency classifier trained on one domain transfers effectively to others without retraining. The code is available at: https://github.com/Applied-MachineLearning-Lab/ACL2026_Tandem.