Prior studies have shown that distinguishing text generated by Large Language Models (LLMs) from human-written one is highly challenging for humans, and often no better than random guessing. To verify the generalizability of this finding across languages and domains, we perform an extensive case study to identify the upper bound of human detection accuracy. Across 16 datasets covering 9 languages and 9 domains, 19 annotators achieved an average detection accuracy of 87.6%, thus challenging previous conclusions. We find that major gaps between human and machine text lie in concreteness, cultural nuances, and diversity. Prompting by explicitly explaining the distinctions in the prompts can partially bridge the gaps in over 50% of the cases. However, we also find that humans do not always prefer human-written text, particularly when they cannot clearly identify its source. We release our dataset, the human labels, and the annotator metadata at https://github.com/xnlp-lab/HumanEval-MGT.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
The widespread deployment of large language models (LLMs) makes detecting LLM-Generated text a critical security task. Existing methods, primarily relying on output probabilities from proxy models or single semantic features, suffer from distribution misalignment and limited interpretability. We observe that machine-generated text exhibits a directionally consistent systematic translation relative to human-written text within the joint semantic-structural space. Accordingly, we propose ProSSD, a statistical framework utilizing supervised subspace learning to extract compact features and construct conditional semantic distributions based on syntactic structures. By employing a likelihood ratio test, we derive a modified Mahalanobis distance, weighted by the Wasserstein distance, as the discriminative metric. Experiments demonstrate ProSSD’s superior robustness and computational efficiency across cross-domain, cross-model, and adversarial scenarios. Furthermore, we reveal the phenomena of systematic semantic translation and semantic collapse in machine-generated text, offering interpretable statistical insights into LLM generation behaviors.
From Answers to Arguments: Toward Trustworthy Clinical Diagnostic Reasoning with Toulmin-Guided Curriculum Goal-Conditioned Learning
PDF ↗The integration of Large Language Models (LLMs) into clinical decision support is critically obstructed by their opaque and often unreliable reasoning. In the high-stakes domain of healthcare, correct answers alone are insufficient; clinical practice demands full transparency to ensure patient safety and enable professional accountability. A pervasive and dangerous weakness of current LLMs is their tendency to produce "correct answers through flawed reasoning." This issue is far more than a minor academic flaw; such process errors signal a fundamental lack of robust understanding, making the model prone to broader hallucinations and unpredictable failures when faced with real-world clinical complexity. In this paper, we establish a framework for trustworthy clinical argumentation by adapting the Toulmin model to the diagnostic process. We propose a novel training pipeline: Curriculum Goal-Conditioned Learning (CGCL), designed to progressively train LLM to generate diagnostic arguments that explicitly follow this Toulmin structure. CGCL’s progressive three-stage curriculum systematically builds a solid clinical argument: (1) extracting facts and generating differential diagnoses; (2) justifying a core hypothesis while rebutting alternatives; and (3) synthesizing the analysis into a final, qualified conclusion. We validate CGCL using T-Eval, a quantitative framework measuring the integrity of the diagnosis reasoning. Experiments show that our method achieves diagnostic accuracy and reasoning quality comparable to resource-intensive Reinforcement Learning (RL) methods, while offering a more stable and efficient training pipeline.
Omni-RewardBench: Toward a Comprehensive Evaluation of Generative Reward Models Across Modalities
PDF ↗The rise of Omni-modality Large Language Models (OLLMs) capable of jointly processing text, audio, and visual inputs marks a major step toward general intelligence. Ensuring their alignment with human preferences requires effective Omni-modality Reward Models (ORMs), which serve as surrogates for human judgment to guide OLLMs behavior. However, ORMs evaluation remains underdeveloped in the previous literature. Existing benchmarks are largely text-centric or limited to bimodal tasks, restricting comprehensive assessment for ORMs. To bridge this gap, we introduce Omni-RewardBench, the first benchmark for comprehensive evaluation of ORMs across modalities. In short, our contributions are threefold: (1) a hybrid automatic-annotation and human-verification pipeline to construct high-quality evaluation data; (2) extensive experiments on 20+ models, including inherently omni-modal and modality-bridged systems. Our experimental results demonstrate that current OLLMs fall short as reward models, revealing several common failure modes such as perception failure, modality dominance failure, and cross-modal fusion failure; and (3) strong correlations between Omni-RewardBench scores and downstream performance (IID r = 0.94, OOD r = 0.72), validating its reliability as a predictor of real-world capability and alignment quality.
Illusions of the Gold Standard: A Large-scale Analysis of Human Evaluation Protocols for Long-form Text Generation
PDF ↗Human evaluation plays a critical role in assessing the quality of generated text. However, the reliability and reproducibility of these evaluations depend on transparent and well-documented protocols—details that are frequently missing in current practice. In this work, we conduct a large-scale analysis of human evaluation protocols for evaluating long-form generation tasks in *CL conference publications from 2023–2025, including a full manual review of 284 papers and LLM-assisted analysis for another 1.8k+ papers. We define a set of 20 reportable criteria related to reproducibility of human evaluation studies, and apply these criteria to systematically examine reporting norms and practices within the community. We find widespread under-reporting of important aspects of human evaluation study design, leading to ambiguity about what was measured and how, who contributed judgments, and how judgments should be interpreted. Based on these findings, we outline actionable recommendations to support more transparent and reproducible reporting in future research.
How Tokenization Limits Phonological Knowledge Representation in Language Models and How to Improve Them
PDF ↗Tokenization is the first step in every language model (LM), yet it never takes the sounds of words into account. We investigate how tokenization influences text-only LMs’ ability to represent phonological knowledge. Through a series of probing experiments, we show that subword-based tokenization systematically weakens the encoding of both local (e.g., rhyme) and global (e.g., syllabification) phonological features. To quantify this effect, we introduce the syllabification-tokenization alignment distance (STAD), a metric that measures the misalignment between a model’s tokenization and the natural syllable boundaries of words, and find that higher misalignment correlates with poorer phonological representations, providing a simple diagnostic for phonology-aware tokenization. To address these limitations, we propose a lightweight IPA-based fine-tuning method that infuses phonological awareness into LMs, leading to consistent improvements across three phonology-related tasks while largely preserving math and general reasoning ability, with 1.1% and 0.9% drops on GSM8K and MMLU, respectively.[Our code is available at <https://github.com/liaodisen/Tokenization-Phonology>]
With the growing potential of large language models (LLMs) in the legal domain, domain-specific finetuning and retrieval-augmented generation (RAG) methods have received widespread attention. However, current methods still suffer from hallucination risk and failing to resolve semantic drift and adapt to varying citation numbers. To address this, we propose Authoritative and Accurate Lawyer (AALawyer), a complementary dual-retriever framework based on the Legal Syllogism and the nature of different legal data. First, we introduce Symbolic Constrained Retrieval (SCR) for closed-set article retrieval, by constraining retrieval to the generative prediction. Second, we build Analogical Precedent Retrieval (APR) to retrieve open-set judicial precedents for reasoning with a newly collected large criminal dataset.Extensive experiments, including LawBench, our Hallucination Risk-Benchmark, and comprehensive ablation studies, demonstrate the effectiveness of AALawyer, which mitigates hallucinations while improving the explainability of legal reasoning.
Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, methods such as GRPO and DAPO suffer from substantial computational cost, since they rely on sampling many rollouts for each prompt. Moreover, in RLVR the relative advantage is often sparse: many samples become nearly all-correct or all-incorrect, yielding low within-group reward variance and thus weak learning signals. In this paper, we introduce ARRoL (**A**ccelerating **R**LV**R** via **o**nline Ro**L**lout Pruning), an online rollout pruning method that prunes rollouts during generation while explicitly steering the surviving ones more correctness-balanced to enhance learning signals. Specifically, ARRoL trains a lightweight quality head on-the-fly to predict the success probability of partial rollouts and uses it to make early pruning decisions. The learned quality head can further weigh candidates to improve inference accuracy during test-time voting. To improve efficiency, we present a system design that prunes rollouts inside the inference engine and re-batches the remaining ones for log-probability computation and policy updates. Across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models (1B-8B), ARRoL improves average accuracy by +2.30 to +2.99 while achieving up to 1.7× training speedup, and yielding up to +8.33 additional gains in average accuracy in test-time voting.
Code LLMs still struggle with code execution reasoning, especially in smaller models. Existing methods rely on supervised fine-tuning (SFT) with teacher-generated explanations, primarily in two forms: (1) input–output (I/O) prediction chains and (2) natural-language descriptions of execution traces. However, intermediate execution steps cannot be explicitly verified during SFT, so the training objective can reduce to merely matching teacher explanations. Moreover, training data is typically collected without explicit control over task difficulty. We introduce ExecVerify, which goes beyond text imitation by incorporating verifiable white-box rewards derived from execution traces, including next-statement prediction and variable value/type prediction. Our work first builds a dataset with multiple difficulty levels via constraint-based program synthesis. Then, we apply reinforcement learning (RL) to reward correct answers about both intermediate execution steps and final outputs, aligning the training objective with semantic correctness at each execution step. Finally, we adopt a two-stage training pipeline that first enhances execution reasoning and then transfers to code generation. Experiments demonstrate that a 7B model trained with ExecVerify achieves performance comparable to 32B models on code reasoning benchmarks and improves pass@1 by up to 5.9% on code generation tasks over strong post-training baselines.
Closing the Spatial Execution Gap in Digital Whiteboards via Verifiable Reinforcement Learning
PDF ↗While multi-modal large language models such as GPT-5 demonstrate exceptional general understanding, they suffer from a fundamental Spatial Execution Gap, failing to translate visual semantics into precise, schema-valid coordinate operations in interactive environments. In this work, we show that model scale alone cannot close this gap; instead, verifiable structured reasoning provides the key to spatial precision. We present a comprehensive pipeline that leverages Group Relative Policy Optimization to enforce a strict Identify-Reason-Verify protocol, effectively shifting the computational burden from parameters to test-time reasoning. By utilizing a multi-agent system to distill optimal reasoning schemas and training on execution-verifiable rewards, our specialized 3B agent achieves 100% format coherence and 81.12% operation accuracy on digital whiteboard tasks. Crucially, our approach outperforms a state-of-the-art frontier model, GPT-5, by 16.75% in operation accuracy. The results suggest that for complex user interface manipulation, small, RL-aligned models with dedicated reasoning protocols are superior to generalist frontier models, offering a promising direction for building reliable web agents.
Backchannels (e.g., ‘yeah’, ‘mhm’, and ‘right’) are short, non-interruptive feedback signals whose lexical form and prosody jointly convey pragmatic meaning. While prior computational research has largely focused on predicting backchannel timing, the relationship between lexico-prosodic form and meaning remains underexplored. We propose a two-stage framework: first, fine-tuning large language models on dialogue transcripts to derive rich contextual representations; and second, learning a joint embedding space for dialogue contexts and backchannel realizations. We evaluate alignment with human perception via triadic similarity judgments (prosodic and cross-lexical) and a context–backchannel suitability task. Our results demonstrate that the learned projections substantially improve context-backchannel retrieval compared to previous methods. In addition, they reveal that backchannel form is highly sensitive to extended conversational context and that the learned embeddings align more closely with human judgments than raw WavLM features.
Output diversity is crucial for Large Language Models as it underpins pluralism and creativity.In this work, we reveal that controlling the language used during model thinking—the *language of thought*—provides a novel and structural source of output diversity.Our preliminary study shows that different thinking languages occupy distinct regions in a model’s thinking space.Based on this observation, we study two repeated sampling strategies under multilingual thinking—*Single-Language Sampling* and *Mixed-Language Sampling*—and conduct diversity evaluation on outputs that are controlled to be in English, regardless of the thinking language used.Across extensive experiments, we demonstrate that switching the thinking language from English to non-English languages consistently increases output diversity, with a clear and consistent positive correlation such that languages farther from English in the thinking space yield larger gains.We further show that aggregating samples across multiple thinking languages yields additional improvements through compositional effects, and that scaling sampling with linguistic heterogeneity expands the model’s diversity ceiling.Finally, we show that these findings translate into practical benefits in pluralistic alignment scenarios, leading to broader coverage of cultural knowledge and value orientations in LLM outputs. Our code is publicly available at https://github.com/iNLP-Lab/Multilingual-LoT-Diversity.
Large Language Models (LLMs) are increasingly integrated into various domains, making knowledge editing techniques crucial yet potentially hazardous. Current editing methods primarily target atomic facts, overlooking the significant risks associated with manipulating “factual opinions”, e.g., documented stances of public figures on societal issues. Such manipulation could reshape public images, influence elections, and alter societal views. To systematically assess this threat, we introduce the Factual Opinion Editing with Evidence (FOE) benchmark, which encompasses 261 public figures, 19 issue categories, and 2,178 complete opinion records. Our evaluations demonstrate that current editing techniques struggle significantly with factual opinions, often achieving only superficial changes while failing to preserve consistency between the edited opinion and the supporting evidence generated by the model. To address this limitation, we further propose a simple yet effective Self-Generated Evidence-Aligned method that achieves opinion–evidence alignment without relying on explicit instructions. Together, our benchmark and method provide a foundation for understanding the emerging security implications of factual opinion editing in LLMs.
Chatbot service providers (e.g., OpenAI) rely on tiered subscription plans to generate revenue, offering black-box access to basic models for free users and advanced models to paying subscribers. However, this approach is unprofitable and inflexible. A pay-to-unlock scheme for premium features (e.g., math, coding) offers a more sustainable alternative. Enabling such a scheme requires a feature-locking technique (FLoTE) that is (i) *effective* in refusing locked features, (ii) *utility-preserving* for unlocked features, (iii) *robust* against evasion or unauthorized credential sharing, and (iv) *scalable* to multiple features and clients. Existing FLoTEs (e.g., password-locked models) fail to meet these criteria. To fill this gap, we present Locket, a more *robust and scalable* FLoTE to enable pay-to-unlock schemes. We develop a framework for adversarial training and merging of feature-locking *adapters*, which enables Locket to selectively disable specific features of a model. Evaluation shows that Locket is effective (100% refusal rate), utility-preserving (\leq 7% utility degradation), robust (\leq 5% attack success rate), and scalable to multiple features and clients.
Long-term memory is a critical challenge for Large Language Model agents, as fixed context windows cannot preserve coherence across extended interactions. Existing memory systems represent conversation history as unstructured embedding vectors, retrieving information through semantic similarity. This paradigm fails to capture the associative structure of human memory, wherein related experiences progressively strengthen interconnections through repeated co-activation. Inspired by cognitive neuroscience, we identify three mechanisms central to biological memory: association, consolidation, and spreading activation, which remain largely absent in current research. To bridge this gap, we propose HeLa-Mem, a bio-inspired memory architecture that models memory as a dynamic graph with Hebbian learning dynamics. HeLa-Mem employs a dual-level organization: (1) an episodic memory graph that evolves through co-activation patterns, and (2) a semantic memory store populated via Hebbian Distillation, wherein a Reflective Agent identifies densely connected memory hubs and distills them into structured, reusable semantic knowledge. This dual-path design leverages both semantic similarity and learned associations, mirroring the episodic-semantic distinction in human cognition. Experiments on LoCoMo demonstrate superior performance across four question categories while using significantly fewer context tokens. Code is available on GitHub: https://github.com/ReinerBRO/HeLa-Mem
LLMs have been widely used in planning, either as planners to generate action sequences end-to-end, or as formalizers to represent the planning domain and problem in a formal language that can derive plans deterministically. However, both lines of work rely on standard benchmarks that include only generic and simplistic environmental specifications, leading to potential overestimation of the planning ability of LLMs and safety concerns in downstream tasks. We bridge this gap by augmenting widely used planning benchmarks with manually annotated, fine-grained, and rich natural language constraints spanning four formally defined categories. Over 4 state-of-the-art reasoning LLMs, 4 formal languages, and 4 datasets, we show that the introduction of one-sentence constraints consistently halves performance, indicating current LLMs’ lack of robustness and an avenue for future research.
Agents built on vision-language models increasingly face tasks that demand anticipating future states rather than relying on short-horizon reasoning. Generative world models offer a promising remedy: agents could use them as external simulators to foresee outcomes before acting. This paper empirically examines whether current agents can leverage such world models as tools to enhance their cognition. Across diverse agentic and visual question answering tasks, we observe that some agents rarely invoke simulation (fewer than 1%), frequently misuse predicted rollouts (approximately 15%), and often exhibit inconsistent or even degraded performance (up to 5%) when simulation is available or enforced. Attribution analysis further indicates that the primary bottleneck lies in the agents’ capacity to decide when to simulate, how to interpret predicted outcomes, and how to integrate foresight into downstream reasoning. These findings underscore the need for mechanisms that foster calibrated, strategic interaction with world models, paving the way toward more reliable anticipatory cognition in future agent systems.
Why Multimodal In-Context Learning Lags Behind? Unveiling the Inner Mechanisms and Bottlenecks
PDF ↗In-context learning (ICL) enables models to adapt to new tasks via inference-time demonstrations. Despite its success in large language models, the extension of ICL to multimodal settings remains poorly understood in terms of its internal mechanisms and how it differs from text-only ICL. In this work, we conduct a systematic analysis of ICL in multimodal large language models. Using identical task formulations across modalities, we show that multimodal ICL performs comparably to text-only ICL in zero-shot settings but degrades significantly under few-shot demonstrations. To understand this gap, we decompose multimodal ICL into task mapping construction and task mapping transfer, and analyze how models establish cross-modal task mappings, and transfer them to query samples across layers. Our analysis reveals that current models lack reasoning-level alignment between visual and textual representations, and fail to reliably transfer learned task mappings to queries. Guided by these findings, we further propose a simple inference-stage enhancement method that reinforces task mapping transfer. Our results provide new insights into the mechanisms and limitations of multimodal ICL and suggest directions for more effective multimodal adaptation.
While Unified Multimodal Models (UMMs) have achieved remarkable success in cross-modal comprehension, a significant gap persists in their ability to leverage such internal knowledge for high-quality generation. We formalize this discrepancy as Conduction Aphasia, a phenomenon where models accurately interpret multimodal inputs but struggle to translate that understanding into faithful and controllable synthesis. To address this, we propose UniCorn, a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. By partitioning a single UMM into three collaborative roles: Proposer, Solver, and Judge, UniCorn generates high-quality interactions via self-play and employs cognitive pattern reconstruction to distill latent understanding into explicit generative signals. To validate the restoration of multimodal coherence, we introduce UniCycle, a cycle-consistency benchmark based on a Text to Image to Text reconstruction loop. Extensive experiments demonstrate that UniCorn achieves comprehensive and substantial improvements over the base model across six general image generation benchmarks. Notably, it achieves SOTA performance on TIIF(73.8), DPG(86.8), CompBench(88.5), and UniCycle while further delivering substantial gains of +5.0 on WISE and +6.5 on OneIG. These results highlight that our method significantly enhances T2I generation while maintaining robust comprehension, demonstrating the scalability of fully self-supervised refinement for unified multimodal intelligence.
Unleashing Spatial Reasoning in Multimodal Large Language Models via Textual Representation Guided Reasoning
PDF ↗Existing Multimodal Large Language Models (MLLMs) struggle with 3D spatial reasoning, as they fail to construct structured abstractions of the 3D environment depicted in video inputs. To bridge this gap, drawing inspiration from cognitive theories of allocentric spatial reasoning, we investigate how to enable MLLMs to model and reason over text-based spatial representations of video. Specifically, we introduce Textual Representation of Allocentric Context from Egocentric Video (TRACE), a prompting method that induces MLLMs to generate text-based representations of 3D environments as intermediate reasoning traces for more accurate spatial question answering. TRACE encodes meta-context, camera trajectories, and detailed object entities to support structured spatial reasoning over egocentric videos. Extensive experiments on VSI-Bench and OST-Bench demonstrate that TRACE yields notable and consistent improvements over prior prompting strategies across a diverse range of MLLM backbones, spanning different parameter scales and training schemas. We further present ablation studies to validate our design choices, along with detailed analyses that probe the bottlenecks of 3D spatial reasoning in MLLMs.