Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99

Long chain-of-thought reasoning has made autoregressive decoding the dominant inference cost of modern large language models. Existing methods target either the input side (latent compression) or the output side (speculative decoding and multi-token prediction, MTP), but the two lines of work have been pursued independently. Moreover, output-side methods must incur an expensive verifier pass to validate the unreliable draft tokens predicted by MTP. To address these issues, we propose Pair-In, Pair-Out (PIPO), which unifies both sides by viewing a latent compressor and an MTP head as mirror-image operations: the compressor folds two input tokens into one latent representation, while the MTP head unfolds one hidden state into one additional output token. To remove the verifier cost without sacrificing reliability, PIPO trains a lightweight confidence head that decides whether draft tokens should be accepted. We observe that On-Policy Distillation (OPD) naturally matches the rejection-sampling criterion of speculative decoding, so the confidence head can be trained alongside OPD with negligible extra cost. Experiments on AIME 2025, GPQA-Diamond, LiveCodeBench v6, and LongBench v2 with Qwen3.5-4B and 9B backbones show that PIPO improves pass@4 over regular decoding by up to +7.15 points, while delivering up to 2.64× first-token-latency and 2.07× per-token-latency speedups.

Hybrid computer-use agents can act through screenshots or call text tools. We find that having a tool available does not settle which way the effect goes. Under one identical GUI-MCP harness on the OSWorld-MCP benchmark (309 tasks), the same MCP tools improve a reasoning model by +4.0pp and degrade a non-reasoning model by -5.9pp (5 runs each, both beyond 2 SE). What separates the two is tool-decision behavior. The non-reasoning policy ignores, misnames, or falsely terminates around tools. The reasoning model avoids these failures, yet still calls a tool on only 55/309 tasks, 23.9

Large language models are increasingly deployed for personalized interaction, and demographic conditioning via user profiles is a widely adopted strategy for cultural adaptation. We ask whether this approach genuinely serves individual users or achieves accuracy by erasing individual distinctiveness. Studying seven models including frontier GPT-5.1 on the World Values Survey, we find that demographic profiles improve value alignment accuracy for most models, but at a systematic cost to individuality. That is, models pull responses toward demographic group centroids rather than preserving individual differences, a behavioral pattern we term alignment by stereotyping. Permutation tests (10,000 permutations, six demographic attributes, seven models) certify that top-performing models compress individuals far above the human baseline; within-family scaling amplifies this tradeoff while degrading intrinsic cultural understanding. Using a synthetic dialogue dataset validated on real human-chatbot conversations from PRISM (Kirk et al., 2024), we further show that distributing demographic signals across conversational turns partially suppresses prototype retrieval compared to compact demographic labels, a finding validated on real conversations via PRISM but requiring replication at larger scale.

Singular Value Decomposition (SVD) enables hardware-agnostic LLM compression via low-rank approximation, yet optimal rank allocation remains a bottleneck. Existing methods predominantly derive layer importance from performance-oriented proxies. Yet, these metrics fail to distinguish between representational importance and structural compressibility, consequently obscuring the fine-grained influence of spectral distribution shape. We demonstrate this disconnect through spectral analysis, revealing that layers with similar information capacity can exhibit markedly different singular value decay behaviors, corresponding to varying degrees of redundancy in the spectral tail. This imperfect coupling implies that allocation strategies driven solely by importance leave significant compression opportunities underexploited. To address this gap, we propose HiSVD, a hierarchical rank allocation framework with two stages: (1) Capacity-Anchored Baseline Allocation, which preserves representational stability by aligning rank budgets with information capacity; and (2) Redundancy-Aware Refinement, which modulates this baseline using tail redundancy to penalize structural excess. Experiments on LLMs demonstrate that HiSVD achieves superior compression efficiency, significantly outperforming state-of-the-art baselines by effectively exploiting this spectral heterogeneity.

Large reasoning models achieve high accuracy through extended chain-of-thought but generate 5–8 more tokens than necessary, applying verbose reasoning uniformly regardless of problem difficulty. We propose Hint Tuning, a data-efficient approach that teaches models to calibrate reasoning depth. Our key insight: the corresponding instruct model serves as an ideal difficulty probe. By testing what the instruct model can solve with varying guidance, we automatically construct training data across three states: No-Hint (direct answer), Sparse-Hint (minimal prefix), and Full-Hint (complete reasoning). This converts the abstract challenge of difficulty labeling into a measurable consistency check between the instruct and reasoning models. With only 1K self-annotated samples, Hint Tuning achieves 24–66

Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. Unlike traditional model compression, which needs retraining, recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. In this paper, we focus on the dynamic depth of LLM generation. A token-position aware layer skipping framework is proposed to save 1.5x times operations efficiently while maintaining performance. We first observed that tokens predicted later have lower perplexity and thus require less computation. Then, we propose a training-free algorithm called Position-Aware Depth Decay Decoding (D^3), which leverages a power-law decay function, ⌊ L × (α^i) ⌋, to determine the number of layers to retain when generating token T_i. Remarkably, without any retraining, the D^3 achieves success across a wide range of generation tasks for the first time. Experiments on large language models (the Llama) with 7 ∼ 70 billion parameters show that D^3 can achieve an average 1.5x speedup compared with the full-inference pipeline while maintaining comparable performance with nearly no performance drop (<1%) on the GSM8K and BBH benchmarks.

Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, yet they often struggle with context-faithfulness generations that properly reflect contextual knowledge. While existing approaches focus on enhancing the decoding strategies, they ignore the fundamental mechanism of how contextual information is processed within LLMs' internal states. As a result, LLMs remain limited in their ability to fully leverage contextual knowledge. In this paper, we propose Context-aware Layer Enhancement (CaLE), a novel intervention method that enhances the utilization of contextual knowledge within LLMs' internal representations. By employing V-usable information analysis, CaLE strategically amplifies the growth of contextual information at an optimal layer, thereby enriching representations in the final layer. Our experiments demonstrate that CaLE effectively improves context-faithful generation in Question-Answering tasks, particularly in scenarios involving unknown or conflicting contextual knowledge.

Effective engagement by large language models (LLMs) requires adapting responses to users' sociodemographic characteristics, such as age, occupation, and education level. While many real-world applications leverage dialogue history for contextualization, existing evaluations of LLMs' behavioral adaptation often focus on single-turn prompts. In this paper, we propose a framework to evaluate LLM adaptation when attributes are introduced either (1) explicitly via user profiles in the prompt or (2) implicitly through multi-turn dialogue history. We assess the consistency of model behavior across these modalities. Using a multi-agent pipeline, we construct a synthetic dataset pairing dialogue histories with distinct user profiles and employ questions from the Value Survey Module (VSM 2013) (Hofstede and Hofstede, 2016) to probe value expression. Our findings indicate that most models adjust their expressed values in response to demographic changes, particularly in age and education level, but consistency varies. Models with stronger reasoning capabilities demonstrate greater alignment, indicating the importance of reasoning in robust sociodemographic adaptation.

Recent thinking models trained with reinforcement learning and backward-checking CoT often suffer from overthinking: they produce excessively long outputs even on simple problems, wasting computation. Existing evaluations, based on token efficiency, give an incomplete view as they neglect problem difficulty and intermediate computation costs. We formalize reasoning efficiency as a relative measure between thinking and instruct models, treating instruct models as the minimal-effort baseline. A systematic study across four thinking models and multiple benchmarks reveals two consistent patterns: (i) instruct models achieve higher efficiency overall, and (ii) problem difficulty affects efficiency, with thinking models wasting computation on easy problems but providing value on harder ones. Building on this insight, we propose COTHINK, a simple two-stage pipeline: an instruct model drafts a brief outline, and a thinking model expands it. On GSM8K, MATH500, and AIME24, COTHINK cuts token usage by 21.1

Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. However, not all requests posed to LLMs are equally difficult to handle. Through analysis, we show that for some tasks, LLMs can achieve results comparable to the final output at some intermediate layers. That is, not all layers of LLMs are necessary during inference. If we can predict at which layer the inferred results match the final results (produced by evaluating all layers), we could significantly reduce the inference cost. To this end, we propose a simple yet effective algorithm named AdaInfer to adaptively terminate the inference process for an input instance. AdaInfer relies on easily obtainable statistical features and classic classifiers like SVM. Experiments on well-known LLMs like the Llama2 series and OPT, show that AdaInfer can achieve an average of 17.8% pruning ratio, and up to 43% on sentiment tasks, with nearly no performance drop (<1%). Because AdaInfer does not alter LLM parameters, the LLMs incorporated with AdaInfer maintain generalizability across tasks.

Towards energy-efficient artificial intelligence similar to the human brain, the bio-inspired spiking neural networks (SNNs) have advantages of biological plausibility, event-driven sparsity, and binary activation. Recently, large-scale language models exhibit promising generalization capability, making it a valuable issue to explore more general spike-driven models. However, the binary spikes in existing SNNs fail to encode adequate semantic information, placing technological challenges for generalization. This work proposes the first fully spiking mechanism for general language tasks, including both discriminative and generative ones. Different from previous spikes with 0,1 levels, we propose a more general spike formulation with bi-directional, elastic amplitude, and elastic frequency encoding, while still maintaining the addition nature of SNNs. In a single time step, the spike is enhanced by direction and amplitude information; in spike frequency, a strategy to control spike firing rate is well designed. We plug this elastic bi-spiking mechanism in language modeling, named SpikeLM. It is the first time to handle general language tasks with fully spike-driven models, which achieve much higher accuracy than previously possible. SpikeLM also greatly bridges the performance gap between SNNs and ANNs in language modeling. Our code is available at https://github.com/Xingrun-Xing/SpikeLM.

Supervised fine-tuning (SFT) is crucial for adapting Large Language Models (LLMs) to specific tasks. In this work, we demonstrate that the order of training data can lead to significant training imbalances, potentially resulting in performance degradation. Consequently, we propose to mitigate this imbalance by merging SFT models fine-tuned with different data orders, thereby enhancing the overall effectiveness of SFT. Additionally, we introduce a novel technique, "parameter-selection merging," which outperforms traditional weighted-average methods on five datasets. Further, through analysis and ablation studies, we validate the effectiveness of our method and identify the sources of performance improvements.

High computational cost, data collection, and difficulty in distributed training are the three significant barriers in pre-training large language models (LLMs) for many researchers. In this paper, we try to solve the question ''Under constrained computational resources, what type of model design(eg. model size, model architecture) should I train in order to to achieve the best possible performance?" To answer this question, based on Scaling Laws for LLM, we introduce nanoLM: an affordable LLM Study Benchmark via Accurate Loss Prediction across scales. This benchmark unlocks a new LLM study paradigm without direct training. Under the loss basin area, the training loss and model size can be accurately fitted as a power law. This allows us to extrapolate LM from small- to large-scale. For example, with just 13.1%, 14.2% of the total pretraining cost, we can accurately forecast the loss for models sized 26B and 52B. To ensure compatibility with mainstream Transformer architectures, nanoLM offers support for decoder-only structures (eg., GPT), encoder-only structures (eg., BERT), and encoder-decoder structures (eg., T5). Considering that excessive model parameters might lead to GPU memory overflow, nanoLM also supports for data parallelism strategies. Our goal with nanoLM is to empower researchers to make cheap and meaningful comparisons of varying model designs at large scales. We also aspire for our benchmark to serve as a bridge between the academic community and the industry.

Large language models (LLMs) are considered important approaches towards foundational machine intelligence, achieving remarkable success in Natural Language Processing and multimodal tasks, among others. However, the carbon footprints and financial costs originating from heavy pre-training computation is a non-negligible issue. Progressive training methods, inspired by the neurogenesis process that grows neural structures, have shown potential to accelerate LLM pre-training. However, the algorithms, implementation, and practices for progressively training LLMs beyond 100B parameters remain underexplored. In this paper, we show that our model, namely FLM-101B, trained with our growth strategy under a budget of $100K, reaches 80% of the baselines' performances with only 10% of their floating-point operations. We believe that further studies on progressive training will benefit the community by cutting down the costs and promoting green AI. The checkpoint of FLM-101B is released at https://huggingface.co/CofeAI/FLM-101B.

As language models scale up, it becomes increasingly expensive to verify research ideas because conclusions on small models do not trivially transfer to large ones. A possible solution is to establish a generic system that accurately predicts certain metrics for large models without training them. Existing scaling laws require hyperparameter search on the largest models, limiting their predicative capability. In this paper, we present an approach (namely μScaling) to predict the pre-training loss, based on our observations that Maximal Update Parametrization (μP) enables accurate fitting of scaling laws close to common loss basins in hyperparameter space. With μScaling, different model designs can be compared on large scales by training only their smaller counterparts. Further, we introduce nanoLM: an affordable LLM pre-training benchmark that facilitates this new research paradigm. With around 14 forecast the loss for models up to 52B. Our goal with nanoLM is to empower researchers with limited resources to reach meaningful conclusions on large models. We also aspire for our benchmark to serve as a bridge between the academic community and the industry. Code for μScaling is available at https://github.com/cofe-ai/Mu-scaling. Code for nanoLLM will be available later.

Conversation Recommender System (CRS) engage in multi-turn conversations with users and provide recommendations through responses. As user preferences evolve dynamically during the course of the conversation, it is crucial to understand natural interaction utterances to capture the user’s dynamic preference accurately. Existing research has focused on obtaining user preference at the entity level and natural language level, and bridging the semantic gap through techniques such as knowledge augmentation, semantic fusion, and prompt learning. However, the representation of each level remains under-explored. At the entity level, user preference is typically extracted from Knowledge Graphs, while other modal data is often overlooked. At the natural language level, user representation is obtained from a fixed language model, disregarding the relationships between different contexts. In this paper, we propose U ser- a ugmented Conversation Recommendation via M ulti-modal graph learning and C ontext Mining ( UaMC ) to address above limitations. At the entity level, we enrich user preference by leveraging multi-modal knowledge. At the natural language level, we employ contrast learning to extract user preference from similar contexts. By incorporating the enhanced representation of user preference, we utilize prompt learning techniques to generate responses related to recommended items. We conduct experiments on two public CRS benchmarks, demonstrating the effectiveness of our approach in both the recommendation and conversation subtasks.

Empathy is one of the fundamental abilities of dialog systems. In order to build more intelligent dialogue systems, it’s important to learn how to demonstrate empathy toward others. Existing studies focus on identifying and leveraging the user’s coarse emotion to generate empathetic responses. However, human emotion and dialog act (e.g., intent) evolve as the talk goes along in an empathetic dialogue. This leads to the generated responses with very different intents from the human responses. As a result, empathy failure is ultimately caused. Therefore, using fine-grained emotion and intent sequential data on conversational emotions and dialog act is crucial for empathetic response generation. On the other hand, existing empathy models overvalue the empathy of responses while ignoring contextual relevance, which results in repetitive model-generated responses. To address these issues, we propose a Multi-Factor sequence Fusion framework (EmpMFF) based on conditional variational autoencoder. To generate empathetic responses, the proposed EmpMFF encodes a combination of contextual, emotion, and intent information into a continuous latent variable, which is then fed into the decoder. Experiments on the EmpatheticDialogues benchmark dataset demonstrate that EmpMFF exhibits exceptional performance in both automatic and human evaluations.

Information extraction (IE) mainly focuses on three highly correlated subtasks, i.e., entity extraction, relation extraction and event extraction. Recently, there are studies using Abstract Meaning Representation (AMR) to utilize the intrinsic correlations among these three subtasks. AMR based models are capable of building the relationship of arguments. However, they are hard to deal with relations. In addition, the noises of AMR (i.e., tags unrelated to IE tasks, nodes with unconcerned conception, and edge types with complicated hierarchical structures) disturb the decoding processing of IE. As a result, the decoding processing limited by the AMR cannot be worked effectively. To overcome the shortages, we propose an Interactive Information Extraction (InterIE) model based on a novel Semantic Information Graph (SIG). SIG can guide our InterIE model to tackle the three subtasks jointly. Furthermore, the well-designed SIG without noise is capable of enriching entity and event trigger representation, and capturing the edge connection between the information types. Experimental results show that our InterIE achieves state-of-the-art performance on all IE subtasks on the benchmark dataset (i.e., ACE05-E+ and ACE05-E). More importantly, the proposed model is not sensitive to the decoding order, which goes beyond the limitations of AMR based methods.

With the continuous development of natural language processing, R elation e xtraction (RE) has been intensively studied and well performed in extracting relations from unstructured texts in both English and modern Chinese. In this paper, we study to extract relations from a special type of text, that is, Chinese textual description of Han Dynasty Stone Reliefs (HanDSR). We aim to develop an efficient relation extractor for special interests with a small number of samples. The problem is challenging due to the large number of rare words in the texts and the mixed-use of modern and ancient Chinese in the same sentence without a domain corpus. To address these problems, we propose a relation extraction method based on dependency parsing and utilize the information of HanDSR on the basic parser. To exploit the representation of dependency trees, we design five dependency semantic path patterns(DSPPs) to extract relation triples of special interests. Besides, we build the HanDSR Treebank that includes 4190 sentences, 28124 dependency trees, following the annotation format of the Penn Chinese Treebank 8.0, which addresses the lack of domain-specific corpus and could be used in extract relations from such texts. Extensive experiments on HanDSR dataset demonstrate the accuracy and efficiency of our solution. The experimental results illustrate that our proposal significantly outperforms the rule-based relation extraction model in both effectiveness and efficiency.