Consistency models (CMs) have shown promise in the efficient generation of both image and text. This raises the natural question of whether we can learn a unified CM for efficient multimodal generation (e.g., text-to-image) and understanding (e.g., image-to-text). Intuitively, such a model could be acquired by applying the consistency distillation (CD) to existing unified multimodal models. However, the key challenge is establishing a unified denoising perspective for both image and text generation, which is essential for establishing the consistency mapping. To tackle this, at the representation level, we advocate for discrete tokens for both modalities to best preserve language modeling capabilities. Critically, instead of defining the text denoising trajectory via recent discrete diffusion language modeling principles, we specify it using the parallel decoding trace of an autoregressive language model, benefiting from the latter’s superior performance in general text generation tasks. The denoising trajectory of image tokens adheres to standard discrete diffusion. We train our unified consistency models (UniCMs) on these combined multimodal trajectories simultaneously with a unified objective. We introduce a trajectory segmentation strategy to improve the training convergence. Empirically, in text-to-image generation, UniCMs outperform SD3 on GenEval and Image Reward, while requiring only approximately 1/8 of the sampling time. Meanwhile, in image-to-text generation, UniCMs surpass Show-o on the MMMU benchmark while being 1.5 × faster at long-sequence generating speed.

Diffusion large language models (dLLMs) accelerate language generation by predicting multiple masks in a single forward pass.However, existing dLLMs can suffer from unreliable predictions in early denoising stages under aggressive parallelism strategies, leading to errors that can propagate to later stages. To tackle this issue, we present Consistency Forcing (CForce) for dLLMs, a distillation method to force the mask predictions of early stages to align with those of later stages. CForce trains the model on pre-collected self-rollout trajectories, thereby improving training-inference alignment.We introduce Confidence Adaptive KL Divergence as a distillation objective to conjoin the merits of forward and reverse KL.We further provide a theoretical analysis for the consistency objective to explain why CForce can approximately minimize the prediction error of early stages.Critically, the same formulation applies to both mask-to-token decoding and edit-capable decoding; in the edit-capable case, later token-to-token refinements provide additional supervision for earlier masked-state predictions. Experiments on non-edit and edit-capable LLaDA models show improved speed-quality trade-offs, especially under high-parallelism decoding budgets.

Block Diffusion Language Models (BD-LMs) support KV caching and flexible-length generation, but their native Single-Block Diffusion (SingleBD) inference leaves inter-block parallelism unused. We study Multi-Block Diffusion (MultiBD), where a running-set of consecutive blocks is refined concurrently. Existing teacher-forced BD-LMs observe only one noisy block conditioned on a clean prefix, while diffusion forcing exposes a long noisy suffix whose state distribution still differs from practical MultiBD inference. We bridge this gap with Multi-Block Diffusion Language Models (MBD-LMs), obtained by post-training BD-LMs using Multi-block Teacher Forcing (MultiTF). MultiTF trains on bounded noise-groups conditioned on clean prefixes, with heterogeneous slot-wise noise patterns that better match inference. An optimized Block Buffer further preserves prefix-cache reuse and static input shapes, translating inter-block parallelism into wall-clock acceleration. On math and code benchmarks, MBD-LLaDA2-Mini increases average Tokens Per Forward pass (TPF) from 3.47 to 6.19 while maintaining comparable accuracy (80.0% versus 81.0%). Combined with DMax, it reaches 9.34 TPF with a 1.02-point accuracy drop. Project Page: https://sjtu-deng-lab.github.io/mbd-lms

Large language model (LLM) agents have exhibited strong problem-solving competence across domains like research and coding. Yet, it remains underexplored whether LLM agents can tackle compounding real-world problems that require a diverse set of tools to complete. Given a broad, heterogeneous tool repository, LLM agents must not only select appropriate tools based on task planning analysis but also strategically schedule the execution order to ensure efficiency. This paper introduces TPS-Bench to benchmark the ability of LLM agents in solving such problems that demand Tool Planning and Scheduling. TPS-Bench collects 200 compounding tasks of two difficulty levels, based on a tool repository containing hundreds of model context protocol (MCP) tools. In particular, each task is composed of multiple subtasks, such as web search, map navigation, calendar checking, etc., and each subtask can be completed by a basic tool. Our evaluation emphasizes both task completion rate and efficiency. The empirical studies on popular closed-source and open-source LLMs indicate that most models can perform reasonable tool planning, but differ in scheduling. For example, GLM-4.5 achieves an outperforming task completion rate of 64.72% with extensive sequential tool calls, hence suffering from significantly long execution time. By contrast, GPT-4o prioritizes parallel tool calls but achieves only a 45.08% completion rate. Considering reinforcement learning (RL) can be a viable way to improve the scheduling efficiency without compromising performance, we perform an initial study on Qwen3-1.7B and witness a 14% reduction in execution time alongside a 6% gain in task completion rate based on only 597 RL training samples.

This paper identifies that misinterpreting the context can be a significant issue during the reasoning process of large language models, spanning from smaller models like Llama3.2-3B-Instruct to cutting-edge ones like DeepSeek-R1. We introduce a novel, post-training approach called Stick to the Facts (SIFT) to tackle this. SIFT leverages increasing inference-time compute to ground LLM reasoning in contexts. At the core of SIFT lies the Sticker, which is generated by the model itself to explicitly emphasize the key information within the context. Given the Sticker, SIFT generates two predictions—one from the Sticker alone and one from the query augmented with the Sticker. If they differ, the Sticker is sequentially refined via forward optimization (to better align the extracted facts with the query) and inverse generation (to conform with the model’s inherent tendencies) for more faithful reasoning outcomes. Studies across diverse models (from 3B to 100B+) and benchmarks (e.g., MATH, AIME) reveal consistent performance improvements. Notably, SIFT improves the pass@1 accuracy of DeepSeek-R1 on AIME2024 from 78.33% to 85.67% and that on AIME2025 from 69.8% to 77.33%. Code will be public after acceptance.

Large Language Models (LLMs) have achieved impressive performance across a range of natural language processing tasks. However, recent advances demonstrate that further gains—particularly in complex reasoning tasks—require more than merely scaling up model sizes or training data. One promising direction is to enable models to “think” during the reasoning process. Recently, Quiet-STaR significantly improves reasoning by generating token-level thought traces, but incurs substantial inference overhead. In this work, we propose Fast Quiet-STaR, a more efficient reasoning framework that preserves the benefits of token-level reasoning while reducing computational cost. Our method introduces a curriculum-learning-based training strategy that gradually reduces the number of thought tokens, enabling the model to internalize more abstract and concise reasoning processes. We further extend this approach to the standard Next Token Prediction (NTP) setting through reinforcement learning-based fine-tuning, resulting in Fast Quiet-STaR NTP, which eliminates the need for explicit thought token generation during inference. Experiments on four benchmark datasets with Mistral 7B and Qwen2.5 7B demonstrate that Fast Quiet-STaR consistently outperforms Quiet-STaR in terms of average accuracy under the same inference time budget. Notably, Fast Quiet-STaR NTP achieves an average accuracy improvement of 9% on Mistral 7B and 5.7% on Qwen2.5 7B, while maintaining the same inference latency.

Mixture of experts (MoE) has become the standard for constructing production-level large language models (LLMs) due to its promise to boost model capacity without causing significant overheads. Nevertheless, existing MoE methods usually enforce a constant top-k routing for all tokens, which is arguably restrictive because various tokens (e.g., "<EOS>” vs. “apple”) may require various numbers of experts for feature abstraction. Lifting such a constraint can help make the most of limited resources and unleash the potential of the model for downstream tasks. In this sense, we introduce AdaMoE to realize token-adaptive routing for MoE, where different tokens are permitted to select a various number of experts. AdaMoE makes minimal modifications to the vanilla MoE with top-k routing—it simply introduces a fixed number of null experts, which do not consume any FLOPs, to the expert set and increases the value of k. AdaMoE does not force each token to occupy a fixed number of null experts but ensures the average usage of the null experts with a load-balancing loss, leading to an adaptive number of null/true experts used by each token. AdaMoE exhibits a strong resemblance to MoEs with expert choice routing while allowing for trivial auto-regressive modeling. AdaMoE is easy to implement and can be effectively applied to pre-trained (MoE-)LLMs. Extensive studies show that AdaMoE can reduce average expert load (FLOPs) while achieving superior performance. For example, on the ARC-C dataset, applying our method to fine-tuning Mixtral-8x7B can reduce FLOPs by 14.5% while increasing accuracy by 1.69%.Code is available at this link.

The detection of machine-generated text, especially from large language models (LLMs), is crucial in preventing serious social problems resulting from their misuse. Some methods train dedicated detectors on specific datasets but fall short in generalizing to unseen test data, while other zero-shot ones often yield suboptimal performance. Although the recent DetectGPT has shown promising detection performance, it suffers from significant inefficiency issues, as detecting a single candidate requires querying the source LLM with hundreds of its perturbations. This paper aims to bridge this gap. Concretely, we propose to incorporate a Bayesian surrogate model, which allows us to select typical samples based on Bayesian uncertainty and interpolate scores from typical samples to other samples, to improve query efficiency. Empirical results demonstrate that our method significantly outperforms existing approaches under a low query budget. Notably, when detecting the text generated by LLaMA family models, our method with just 2 or 3 queries can outperform DetectGPT with 200 queries.