Reinforcement learning with verifiable rewards (RLVR) has advanced reasoning capabilities in multimodal large language models. However, existing methods typically treat visual inputs as deterministic, overlooking the perceptual ambiguity inherent to the visual modality. Consequently, they fail to distinguish whether a model’s uncertainty stems from complex reasoning or ambiguous perception, preventing the targeted allocation of exploration or learning signals. To address this gap, we introduce DUPL, a dual-uncertainty guided policy learning approach for multimodal RLVR that quantifies and leverages both perceptual uncertainty (via symmetric KL divergence) and output uncertainty (via policy entropy) to guide policy updates. By establishing an uncertainty-driven feedback loop and employing a dynamic branch prioritization mechanism, DUPL recalibrates the policy advantage to focus learning on states with high perceptual or decisional ambiguity, enabling effective targeted exploration beyond passive data augmentation. Evaluated on diverse multimodal reasoning benchmarks spanning mathematical and general domains, DUPL achieves solid gains. It improves Qwen2.5-VL accuracy by up to 12.3% (3B) and 7.9% (7B), and Qwen3-VL-Instruct by up to 10.7% (4B) and 12.4% (8B), consistently outperforming GRPO, while seamlessly generalizing to alternative algorithms (DAPO, +6.5% avg) and architectures (LLaVA-OneVision-1.5, +4.7% avg). These results demonstrate that DUPL is an effective and generalizable approach for multimodal RLVR.

As large language model agents tackle increasingly complex long-horizon tasks, effective post-training becomes critical. Prior work faces fundamental challenges: outcome-only rewards fail to precisely attribute credit to intermediate steps, estimated step-level rewards introduce systematic noise, and Monte Carlo sampling approaches for step reward estimation incur prohibitive computational cost. Inspired by findings that only a small fraction of high-entropy tokens drive effective RL for reasoning, we propose Critical Step Optimization (CSO), which focuses preference learning on verified critical steps—decision points where alternate actions demonstrably flip task outcomes from failure to success. Crucially, our method starts from failed policy trajectories rather than expert demonstrations, directly targeting the policy model’s weaknesses. We use a process reward model (PRM) to identify candidate critical steps, leverage expert models to propose high-quality alternatives, then continue execution from these alternatives using the policy model itself until task completion. Only alternatives that the policy successfully executes to correct outcomes are verified and used as DPO training data, ensuring both quality and policy reachability. This yields fine-grained, verifiable supervision at critical decisions while avoiding trajectory-level coarseness and step-level noise. Experiments on GAIA-Text-103 and XBench-DeepSearch show that CSO achieves 37% and 26% relative improvement over the SFT baseline and substantially outperforms other post-training methods, while requiring supervision at only 16% of trajectory steps. This demonstrates the effectiveness of selective verification-based learning for agent post-training.

Reinforcement learning (RL) has emerged as a powerful framework for improving the reasoning capabilities of large language models (LLMs). However, most existing RL approaches rely on sparse outcome rewards, which fail to credit correct intermediate steps in partially successful solutions. Process reward models (PRMs) offer fine-grained step-level supervision, but their scores are often noisy and difficult to evaluate. As a result, recent PRM benchmarks focus on a more objective capability: detecting the first incorrect step in a reasoning path. However, this evaluation target is misaligned with how PRMs are typically used in RL, where their step-wise scores are treated as raw rewards to maximize. To bridge this gap, we propose Verifiable Prefix Policy Optimization (VPPO), which uses PRMs only to localize the first error during RL. Given an incorrect rollout, VPPO partitions the trajectory into a verified correct prefix and an erroneous suffix based on the first error, rewarding the former while applying targeted penalties only after the detected mistake. This design yields stable, interpretable learning signals and improves credit assignment. Across multiple reasoning benchmarks, VPPO consistently outperforms sparse-reward RL and prior PRM-guided baselines on both Pass@1 and Pass@K.

While large audio language models (LALMs) have driven significant progress in multimodal conversational systems, current benchmarks suffer from critical limitations: they are largely English-centric, use synthetic speech, and fail to provide comprehensive, discriminative evaluation across key dimensions. To fill this gap, we present Voice Chat Bot Bench (VCB Bench), a novel, high-quality Chinese benchmark built exclusively on real human speech. VCB Bench assesses LALMs across three complementary axes: instruction following (including speech-level control beyond text commands), knowledge understanding (including general knowledge, reasoning, and daily dialogue), and robustness (evaluating stability under variations in content, environment, and speaker characteristics). Experiments conducted on representative LALMs reveal notable performance disparities and offer tangible insights for future improvements. VCB Bench serves as a reproducible and fine-grained framework, providing standardized evaluation and practical guidance for the development of Chinese voice conversational models.

Recent advances in Deep Research Agents (DRAs) are transforming automated knowledge discovery and problem-solving.While the majority of existing efforts focus on enhancing policy capabilities via post-training, we propose an alternative paradigm: test-time self-evolving the agent’s ability by iteratively verifying the policy model’s outputs, guided by meticulously crafted rubrics. This approach gives rise to an inference-time scaling of verification, wherein an agent self-improves at test time by evaluating its generated answers to produce iterative feedback and refinements without any additional training. We derive the rubrics based on an automatically constructed DRA Failure Taxonomy, which systematically classifies agent failures into five major categories and thirteen sub-categories. We present DeepVerifier, a rubrics-based outcome reward verifier that leverages the asymmetry of verification and outperforms vanilla agent-as-judge and LLM judge baselines by 12%–48% in meta-evaluation F1 score. To enable practical test-time self-evolution, DeepVerifier integrates as a plug-and-play module during test-time inference. The verifier produces detailed rubric-based feedback, which is fed back to the agent for iterative bootstrapping—refining responses without additional training. This test-time scaling delivers 8%–11% accuracy gains on challenging subsets of GAIA and XBench-DeepResearch when powered by capable closed-source LLMs. Finally, to support open-source advancement, we release DeepVerifier-4K, a curated supervised fine-tuning dataset of 4,646 high-quality agent steps focused on DRA verification. These examples emphasize reflection and self-critique, enabling open models to develop robust verification capabilities.

Although existing model editing methods perform well in recalling exact edit facts, they often struggle in complex scenarios that require deeper semantic understanding rather than mere knowledge regurgitation. Leveraging the strong contextual reasoning abilities of large language models (LLMs), in-context learning (ICL) becomes a promising editing method by comprehending edit information through context encoding. However, this method is constrained by the limited context window of LLMs, leading to degraded performance and efficiency as the number of edits increases. To overcome this limitation, we propose InComeS, a flexible framework that enhances LLMs’ ability to process editing contexts through explicit compression and selection mechanisms. Specifically, InComeS compresses each editing context into the key-value (KV) cache of a special token, enabling efficient handling of multiple edits without being restricted by the model’s context window. Furthermore, specialized cross-attention modules are added to dynamically select the most relevant information from the pool of special tokens, enabling adaptive and effective utilization of edit information. We conduct experiments on diverse model editing benchmarks with various editing formats, and the results demonstrate the effectiveness and efficiency of our method.

With recent advancements in large language models, web agents have been greatly improved. However, dealing with complex and dynamic web environments requires more advanced planning and search abilities. Previous studies usually adopt a greedy one-way search strategy, which may struggle to recover from erroneous states. In this work, we enhance web agents with an explicit rollback mechanism, enabling the agent to revert back to a previous state in its navigation trajectory. This mechanism gives models the flexibility to directly control the search process, leading to an effective and efficient web navigation method. We conduct experiments on two live web navigation benchmarks with zero-shot and fine-tuning settings. The results demonstrate the effectiveness of our proposed approach.

Reinforcement Learning (RL) enhances LLM reasoning, yet a paradox emerges as models scale: strong base models saturate standard benchmarks (e.g., MATH), yielding correct but homogeneous solutions. In such environments, the lack of failure cases causes the advantage signal in group-relative algorithms (e.g., GRPO) to vanish, driving policies into mode collapse. To address this, we propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration. Unlike standard sampling that follows model biases, CUTS flattens the local optimization landscape by sampling uniformly from constrained high-confidence candidates. We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance. Experiments on Qwen3 models demonstrate that our approach prevents policy degeneration and significantly boosts out-of-domain generalization. Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO, validating that maintaining diversity within the semantic manifold is critical for rigorous reasoning.

Graphical User Interface (GUI) grounding requires mapping natural language instructions to precise pixel coordinates. However, due to visually homogeneous elements and dense layouts, models typically grasp semantic intent yet fail to achieve precise localization. While scaling sampling attempts (Pass@k) reveals potential gains, static self-consistency strategies derived from geometric clustering often yield limited improvements, as the model’s predictions tend to be spatially dispersed. In this paper, we propose replacing static consistency strategies with a learnable selection mechanism that selects the optimal target by critiquing its own proposals rendered on the screenshot. Given the significant disparity between the model’s grounding and critiquing capabilities, we propose a co-evolving Propose-and-Critic framework. To jointly optimize these, we introduce a maturity-aware adaptive co-evolutionary reinforcement learning paradigm. This approach dynamically balances the training objectives of proposer and critic, where the diversity of the proposer’s outputs enhances critic robustness, while the critic’s maturing discrimination capability conversely unlocks the proposer’s potential for extensive spatial exploration, fostering the mutual reinforcement and co-evolution of both capabilities, thereby ensuring generalizability to adapt to diverse and complex interface layouts. Extensive experiments over 6 benchmarks show that our method significantly enhances both capabilities.

Assessing the quality of Large Language Model (LLM) outputs becomes especially challenging in high-branching settings, where a single prompt yields many plausible candidates. Existing verifiers typically operate on the surface text (e.g., reward models, LLM judges, majority voting) or on confidence proxies derived from token probabilities, both of which can be brittle: the former can be influenced by stylistic artifacts, while the latter is often miscalibrated. In this paper, we study a third source of information—the model’s hidden states—for binary correctness verification in tasks with a reliable success/failure signal (e.g., deterministic checkers or reference-grounded answers). We find that correct and incorrect solutions exhibit measurable geometric differences in their hidden-state trajectories. To isolate this signal with minimal modeling assumptions, we introduce Clue (Clustering and Experience-based Verification), a training-free, non-parametric verifier. Clue summarizes each reasoning trace by an activation delta—the difference between hidden states at the start and end of the explicit reasoning span—and predicts correctness by comparing this delta to two class centroids computed from labeled experience. Across math (AIME 24/25), scientific QA (GPQA), and a multi-domain benchmark (WebInstruct-verified), Clue improves selection and reranking, with particularly strong gains on smaller or less-calibrated models. For example, on AIME 24 with a 1.5B model, Clue raises accuracy from 56.7% (majority@64) to 70.0% (top-maj@16).

Large Language Models (LLMs) have recently advanced the field of Automated Theorem Proving (ATP), attaining substantial performance gains through widely adopted test-time scaling strategies, notably reflective Chain-of-Thought (CoT) reasoning and increased sampling passes. However, they both introduce significant computational overhead for inference. Moreover, existing cost analyses typically regulate only the number of sampling passes, while neglecting the substantial disparities in sampling costs introduced by different scaling strategies. In this paper, we systematically compare the efficiency of different test-time scaling strategies for ATP models and demonstrate the inefficiency of the current state-of-the-art (SOTA) open-source approaches. We then investigate approaches to significantly reduce token usage and sample passes while maintaining the original performance. Specifically, we propose two complementary methods that can be integrated into a unified EconRL pipeline for amplified benefits: (1) a dynamic Chain-of-Thought (CoT) switching mechanism designed to mitigate unnecessary token consumption, and (2) Diverse parallel-scaled reinforcement learning (RL) with trainable prefixes to enhance pass rates under constrained sampling passes. Experiments on miniF2F and ProofNet demonstrate that our EconProver-GD achieves comparable performance to baseline methods with only 12% of the computational cost. This work provides actionable insights for deploying lightweight ATP models without sacrificing performance.

Reinforcement learning with verifiable rewards (RLVR) has been effective on tasks with structured solutions like math and coding, but its reliance on simple, rule-based verifiers creates a fundamental bottleneck. We find their applicability is surprisingly narrow even in structured domains, a limitation that is compounded at scale: rule-based systems can paradoxically degrade in performance as multi-domain, free-form training data increases. To overcome these challenges, we propose a new RLVR framework that uses a generative verifier to provide soft, probabilistic rewards. Our key insight is that powerful LLMs show high agreement with human evaluators when judging answer correctness given a ground-truth reference, allowing us to automate reward generation without costly human annotation. Our experiments demonstrate the effectiveness of this approach. We show that a compact 7B generative reward model can guide a 7B policy model to decisively outperform models up to 10x its size, including the 72B Qwen2.5-Instruct (by a margin of +8.6%). This effectiveness is robust, holding true across diverse training datasets with answers sourced from experts, web users, and other LLMs, and generalizes strongly to seven out-of-distribution benchmarks. Our work provides a scalable and effective framework for extending RLVR beyond the limitations of pattern-based verification to complex, noisy, real-world domains.

Audio-language pretraining (ALP) holds promise for learning general-purpose audio representation, yet remains underexplored. Crucially, there is no consensus on whether audio–language models can build effective general-purpose audio encoders, nor a systematic understanding of how pretraining objectives behave across diverse tasks and scales.We identify three key barriers: limited scale of audio-text corpora, limited coverage of audio attributes in existing caption corpora, and lack of systematic exploration and evaluation.To fill this gap, we present the first principled empirical study of ALP.We first introduce CaptionStew, a 10.7M caption dataset aggregating open-source audio-text corpora across multiple domains and captioning focuses.We then conduct the first comprehensive evaluation comparing contrastive and captioning objectives for learning audio representation across speech, music, and environmental sound tasks.Our results not only demonstrate that ALP yields competitive, transferable representations, but reveal critical trade-offs: contrastive learning offers superior data efficiency, while captioning exhibits better scalability.Furthermore, we find that the benefits of supervised initialization often diminish at larger scales, challenging common practices.By grounding these claims in empirical evidence, we establish a viable pathway toward general-purpose audio representation learning, guiding future research.

The hallmark of Deep Research agents lies in compositional reasoning, the capacity to aggregate distributed, heterogeneous information into coherent logical insights. However, current agentic systems are often retrieval-heavy but reasoning-light, where success is predominantly determined by simple entity-seeking rather than the multi-step aggregation of scattered evidence. To address this, we propose a data synthesis pipeline WebAggregator, designed to shift the agentic paradigm from retrieval-centric to compositional aggregation. Our approach first employs Proactive Explorer to collect interconnected knowledge, then Compositional Logic Proposer to weave knowledge into complex questions using over 12 composition guidelines derived from a rigorous deconstruction of the Deep Research problem setting. Fine-tuning on this corpus fundamentally transforms agent behavior, fostering deliberate composition reasoning and reduced tool redundancy. The resulting WebAggregator-32B surpasses GPT-4.1 and matches Claude-3.7-Sonnet on GAIA, WebWalkerQA, and XBench. To address the lack of benchmarks that emphasize both reasoning and retrieval, we introduce the WebAggregatorQA testbed, which reveals that even with perfect retrieval, top-tier models still underperformed. These results demonstrate that compositional reasoning, not retrieval, is the true performance ceiling for next-generation research agents.

Recent advancements in proprietary large language models (LLMs), such as those from OpenAI and Anthropic, have led to the development of document reading systems capable of handling raw files with complex layouts, intricate formatting, lengthy content, and multi-modal information. However, the absence of a standardized benchmark hinders objective evaluation of these systems. To address this gap, we introduce DocBench, a benchmark designed to simulate real-world scenarios, where each raw file consists of a document paired with one or more questions. DocBench uniquely evaluates entire document reading systems and adopts a user-centric approach, allowing users to identify the system best suited to their needs.

Large language models (LLMs) often suffer from hallucinations, posing significant challenges for real-world applications. Confidence calibration, as an effective indicator of hallucination, is thus essential to enhance the trustworthiness of LLMs. Prior work mainly focuses on short-form tasks using a single response-level score (macro calibration), which is insufficient for long-form outputs that may contain both accurate and inaccurate claims. In this work, we systematically study atomic calibration, which evaluates factuality calibration at a fine-grained level by decomposing long responses into atomic claims. We further categorize existing confidence elicitation methods into discriminative and generative types, and propose two new confidence fusion strategies to improve calibration. Our experiments demonstrate that LLMs exhibit poorer calibration at the atomic level during long-form generation. More importantly, atomic calibration uncovers insightful patterns regarding the alignment of confidence methods and the changes of confidence throughout generation. This sheds light on future research directions for confidence estimation in long-form generation.

Large Vision-Language Models (LVLMs) have achieved significant progress in tasks like visual question answering and document understanding. However, their potential to comprehend embodied environments and navigate within them remains underexplored. In this work, we first study the challenge of open-vocabulary object navigation by introducing DivScene, a large-scale dataset with 4,614 houses across 81 scene types and 5,707 kinds of target objects. Our dataset provides a much greater diversity of target objects and scene types than existing datasets, enabling a comprehensive task evaluation. We evaluated various methods with LVLMs and LLMs on our dataset and found that current models still fall short of open-vocab object navigation ability. Then, we fine-tuned LVLMs to predict the next action with CoT explanations. We observe that LVLM’s navigation ability can be improved substantially with only BFS-generated shortest paths without any human supervision, surpassing GPT-4o by over 20% in success rates.

Web agents powered by Large Language Models (LLMs) show promise for next-generation AI, but their limited reasoning in uncertain, dynamic web environments hinders robust deployment. In this paper, we identify key reasoning skills essential for effective web agents, i.e., reflection & lookahead, branching, and rollback, and curate trajectory data that exemplifies these abilities by reconstructing the agent’s (inference-time) reasoning algorithms into chain-of-thought rationales. We conduct experiments in the agent self-improving benchmark, OpenWebVoyager, and demonstrate that distilling salient reasoning patterns into the backbone LLM via simple fine-tuning can substantially enhance its performance. Our approach yields significant improvements across multiple benchmarks, including WebVoyager, Mind2web-live, and SimpleQA (web search), highlighting the potential of targeted reasoning skill enhancement for web agents.

LLMs are ideal for decision-making thanks to their ability to reason over long contexts. However, challenges arise when processing speech transcripts that describe complex scenarios, as they are verbose and include repetition, hedging, and vagueness. E.g., during a company’s earnings call, an executive might project a positive revenue outlook to reassure investors, despite uncertainty regarding future earnings. It is crucial for LLMs to incorporate this uncertainty systematically when making decisions. In this paper, we introduce DeFine, a modular framework that constructs probabilistic factor profiles from complex scenarios. It then integrates these profiles with analogical reasoning, leveraging insights from similar past experiences to guide LLMs in making critical decisions in new situations. Our framework separates the tasks of quantifying uncertainty and incorporating it into LLM decision-making. This approach is particularly useful in areas such as consulting and financial deliberation, where making decisions under uncertainty is vital.

The Mixture of Depths (MoD) was introduced to improve computational efficiency by dynamically skipping less important layers, reducing redundant computation while maintaining model capacity. Despite its promise, existing MoD approaches remain under-explored and face two main challenges: (1) high training costs due to the need to train the entire model along with the routers that determine which layers to skip, and (2) performance degradation when important layers are bypassed. In response to the first issue, we propose Router-Tuning, which fine-tunes only the routers on a small dataset, drastically reducing the computational overhead associated with full model training. For the second challenge, we investigate across different architectures and granularities, demonstrating its effectiveness on Attention layers and MoE layers. This method preserves the model’s performance while significantly enhancing computational and memory efficiency. Extensive experiments demonstrate that our approach delivers competitive results while dramatically improving the computation efficiency, e.g., 21% speedup and only a 0.2% performance drop. The code will be released upon acceptance.

GUI agents powered by vision-language models (VLMs) show promise in automating complex digital tasks. However, their effectiveness in real-world applications is often limited by scarce training data and the inherent complexity of these tasks, which frequently require long-tailed knowledge covering rare, unseen scenarios. We propose RAG-GUI , a lightweight VLM that leverages web tutorials at inferencetime. RAG-GUI is first warm-started via supervised finetuning (SFT) and further refined through self-guided rejection sampling fine-tuning (RSF). Designed to be model-agnostic, RAG-GUI functions as a generic plug-in that enhances any VLM-based agent. Evaluatedacross three distinct tasks, it consistently outperforms baseline agents and surpasses other inference baselines by 2.6% to 13.3% acrosstwo model sizes, demonstrating strong generalization and practical plug-and-play capabilities in real-world scenarios.

Agent self-improvement, where agents autonomously train their underlying Large Language Model (LLM) on self-sampled trajectories, shows promising results but often stagnates in web environments due to limited exploration and under-utilization of pretrained web knowledge. To improve the performance of self-improvement, we propose a novel framework that introduces a co-evolving World Model LLM. This world model predicts the next observation based on the current observation and action within the web environment. The World Model serves dual roles: (1) as a virtual web server generating self-instructed training data to continuously refine the agent’s policy, and (2) as an imagination engine during inference, enabling look-ahead simulation to guide action selection for the agent LLM. Experiments in real-world web environments (Mind2Web-Live, WebVoyager, and GAIA-web) show a 10% performance gain over existing self-evolving agents, demonstrating the efficacy and generalizability of our approach, without using any distillation from more powerful close-sourced models.

Mamba’s theoretical infinite-context potential is limited in practice when sequences far exceed training lengths. This work explores unlocking Mamba’s long-context memory ability by a simple-yet-effective method, Recall with Reasoning (RwR), by distilling chain-of-thought (CoT) summarization from a teacher model. Specifically, RwR prepends these summarization as CoT prompts during fine-tuning, teaching Mamba to actively recall and reason over long contexts. Experiments on LONGMEMEVAL and HELMET show that RwR outperforms existing long-term memory methods on the Mamba model. Furthermore, under similar pre-training conditions, RwR improves the long-context performance of Mamba relative to comparable Transformer/hybrid baselines while preserving short-context capabilities, all without changing the architecture.

Large Language Models (LLMs) are prone to hallucination, particularly in long-form generations. A promising direction to mitigate hallucination is to teach LLMs to express uncertainty explicitly when they lack sufficient knowledge. However, existing work lacks direct and fair evaluation of LLMs’ ability to express uncertainty effectively in long-form generation. To address this gap, we first introduce UNCLE, a benchmark designed to evaluate uncertainty expression in both long- and short-form question answering (QA). UNCLE covers five domains and includes more than 1,000 entities, each with paired short- and long-form QA items. Our dataset is the first to directly link short- and long-form QA through aligned questions and gold-standard answers.Along with UNCLE, we propose a suite of new metrics to assess the models’ capabilities to selectively express uncertainty. We then demonstrate that current models fail to convey uncertainty appropriately in long-form generation. We further explore both prompt-based and training-based methods to improve models’ performance, with the training-based methods yielding greater gains. Further analysis of alignment gaps between short- and long-form uncertainty expression highlights promising directions for future research using UNCLE.

Hybrid models that combine state space models (SSMs) with attention mechanisms have demonstrated strong performance by leveraging the efficiency of SSMs and the high recall ability of attention. However, the underlying reasons for these benefits remain insufficiently understood. In this work, we investigate hybrid architectures through the lens of memory utilization and overall performance, and propose a complementary method to further enhance their effectiveness. We focus in particular on the distinction between sequential and parallel integration of SSM and attention layers. Our analysis reveals that sequential hybrids perform better on shorter contexts, whereas parallel hybrids are more effective for longer contexts. Among various configurations, parallel hybrids using a cross-attention to combine SSM and attention outputs perform best. We also introduce a data-centric approach to further improve model performance: continual training on datasets with paraphrases. This method strikes the best balance across various other datasets, enhancing memory recall while preserving other capabilities. It generalizes well across different base models, including pure SSMs, and outperforms architectural modifications aimed at enhancing recall.

While Large Language Models (LLMs) demonstrate impressive capabilities, they still struggle with generating factually incorrect content (i.e., hallucinations). A promising approach to mitigate this issue is enabling models to express uncertainty when unsure. Previous research on uncertainty modeling has primarily focused on short-form QA, but real-world applications often require much longer responses. In this work, we introduce the task of Long-form Generation with Uncertainty (LoGU). We identify two key challenges: Uncertainty Suppression, where models hesitate to express uncertainty, and Uncertainty Misalignment, where models convey uncertainty inaccurately. To tackle these challenges, we propose a refinement-based data collection framework and a two-stage training pipeline. Our framework adopts a divide-and-conquer strategy, refining uncertainty based on atomic claims. The collected data are then used in training through supervised fine-tuning (SFT) and direct preference optimization (DPO) to enhance uncertainty expression. Extensive experiments on three long-form instruction following datasets show that our method significantly improves accuracy, reduces hallucinations, and maintains the comprehensiveness of responses.

Large language models have shown remarkable performance across a wide range of language tasks, owing to their exceptional capabilities in context modeling. The most commonly used method of context modeling is full self-attention, as seen in standard decoder-only Transformers. Although powerful, this method can be inefficient for long sequences and may overlook inherent input structures. To address these problems, an alternative approach is parallel context encoding, which splits the context into sub-pieces and encodes them parallelly. Because parallel patterns are not encountered during training, naively applying parallel encoding leads to performance degradation. However, the underlying reasons and potential mitigations are unclear. In this work, we provide a detailed analysis of this issue and identify that unusually high attention entropy can be a key factor. Furthermore, we adopt two straightforward methods to reduce attention entropy by incorporating attention sinks and selective mechanisms. Experiments on various tasks reveal that these methods effectively lower irregular attention entropy and narrow performance gaps. We hope this study can illuminate ways to enhance context modeling mechanisms.

In this work, we provide an empirical investigation of gist-based context compression methods to improve context processing in large language models. We focus on two key questions: (1) How well can these methods replace full attention models? and (2) What potential failure patterns arise due to compression? Through extensive experiments, we show that while gist-based compression can achieve only slight performance loss on tasks like retrieval-augmented generation and long-document QA, it faces challenges in tasks like synthetic recall. Furthermore, we identify three key failure patterns: lost by the boundary, lost if surprise, and lost along the way. To mitigate these issues, we propose two effective strategies: fine-grained autoencoding, which enhances the reconstruction of original token information, and segment-wise token importance estimation, which adjusts optimization based on token dependencies. Our work provides valuable insights into the understanding of gist token-based context compression and offers practical strategies for improving compression capabilities.

Low-bit quantization improves machine learning model efficiency but surprisingly favors undertrained large language models (LLMs). Larger models or those trained on fewer tokens exhibit less quantization-induced degradation (QiD), while smaller, well-trained models face significant performance losses. To gain deeper insights into this trend, we study over 1500+ quantized LLM checkpoints of various sizes and at different training levels (undertrained or fully trained) in a controlled setting, deriving scaling laws for understanding the relationship between QiD and factors: the number of training tokens, model size and bit width.With our derived scaling laws, we propose a novel perspective that we can use QiD to measure an LLM’s training levels and determine the number of training tokens required for fully training LLMs of various sizes. Moreover, we use the scaling laws to predict the quantization performance of different-sized LLMs trained with tokens. Our projection shows that the low-bit quantization performance of future models, which are expected to be trained with over \textcolor{red}{100~trillion} tokens, may NOT be desirable. This poses a potential challenge for low-bit quantization in the future and highlights the need for awareness of a model’s training level when evaluating low-bit quantization research. To facilitate future research on this problem, we release all the 1500+ quantized checkpoints used in this work at https://huggingface.co/Xu-Ouyang.

The advancement of foundation models has laid the groundwork for building autonomous agents for complex tasks such as web navigation. Recent efforts have also tried to equip the agent with the ability to explore environments and continuously improve over time. However, existing works only focused on building text-only agents in synthetic environments where the reward signals are clearly defined. Such agents can hardly generalize to realistic settings that require multimodal perception ability and provide no ground-truth signal. In this paper, we introduce an innovative multimodal web agent that can autonomously conduct real-world exploration and improve itself. We first train the base model with imitation learning to gain the basic abilities. We then let the agent explore the open web and collect feedback on its trajectories. After that, it further improves its policy by learning from well-performing trajectories judged by another general-purpose model. This exploration-feedback-optimization cycle can continue for several iterations. Experimental results show that our web agent successfully improves itself after each iteration, demonstrating strong performance across multiple test sets. We will release our code and model to encourage future research in this field.

Recent advancements in tree search algorithms guided by verifiers have significantly enhanced the reasoning capabilities of large language models (LLMs), but at the cost of increased computational resources. In this work, we identify two key challenges contributing to this inefficiency: over-exploration due to redundant states with semantically equivalent content, and under-exploration caused by high variance in verifier scoring leading to frequent trajectory switching. To address these issues, we propose FETCH – an e ffici ent tree sear ch framework, which is a flexible, plug-and-play system compatible with various tree search algorithms.Our framework mitigates over-exploration by merging semantically similar states using agglomerative clustering of text embeddings obtained from a fine-tuned SimCSE model. To tackle under-exploration, we enhance verifiers by incorporating temporal difference learning with adjusted 𝜆-returns during training to reduce variance, and employing a verifier ensemble to aggregate scores during inference. Experiments on GSM8K, GSM-Plus, and MATH datasets demonstrate that our methods significantly improve reasoning accuracy and computational efficiency across four different tree search algorithms, paving the way for more practical applications of LLM-based reasoning. The code is available at https://github.com/DeepLearnXMU/Fetch.

We investigate how to elicit compositional generalization capabilities in large language models (LLMs). Compositional generalization empowers LLMs to solve complex problems by combining foundational skills, a critical reasoning ability akin to human intelligence. However, even the most advanced LLMs currently struggle with this form of reasoning. We examine this problem within the framework of in-context learning and find that demonstrating both foundational skills and compositional examples grounded in these skills within the same prompt context is crucial. We refer to this prompt structure as skills-in-context (SKiC). With as few as two exemplars, this in-context learning structure enables LLMs to tackle more challenging problems requiring innovative skill combinations, achieving near-perfect systematic generalization across a broad range of tasks. Intriguingly, SKiC also unlocks the latent potential of LLMs, allowing them to more actively utilize pre-existing internal skills acquired during earlier pretraining stages to solve complex reasoning problems. The SKiC structure is robust across different skill constructions and exemplar choices and demonstrates strong transferability to new tasks. Finally, inspired by our in-context learning study, we show that fine-tuning LLMs with SKiC-style data can elicit zero-shot weak-to-strong generalization, enabling the models to solve much harder problems directly with standard prompting.

Abstract reasoning, the ability to reason from the abstract essence of a problem, serves as a key to generalization in human reasoning. However, eliciting language models to perform reasoning with abstraction remains unexplored. This paper seeks to bridge this gap by introducing a novel structured reasoning format called Abstraction-of-Thought (AoT). The uniqueness of AoT lies in its explicit requirement for varying levels of abstraction within the reasoning process. This approach could elicit language models to first contemplate on the abstract level before incorporating concrete details, which is overlooked by the prevailing step-by-step Chain-of-Thought (CoT) method. To align models with the AoT format, we present AoT Collection, a generic finetuning dataset consisting of 348k high-quality samples with AoT reasoning processes, collected via an automated and scalable pipeline. We finetune a wide range of language models with AoT Collection and conduct extensive evaluations on 23 unseen tasks from the challenging benchmark Big-Bench Hard. Experimental results indicate that models aligned to AoT reasoning format substantially outperform those aligned to CoT in many reasoning tasks.

In this work, we focus on a fundamental yet underexplored problem, event semantic classification in context, to help machines gain a deeper understanding of events. We classify events from six perspectives: modality, affirmation, specificity, telicity, durativity, and kinesis. These properties provide essential cues regarding the occurrence and grounding of events, changes of status that events can bring about, and the connection between events and time. To this end, this paper introduces a novel dataset collected for the semantic classification tasks and several effective models. By incorporating these event properties into downstream tasks, we demonstrate that understanding the fine-grained event semantics benefits downstream event understanding and reasoning via experiments on event extraction, temporal relation extraction, and subevent relation extraction.

One critical issue for chat systems is to stay consistent about preferences, opinions, beliefs and facts of itself, which has been shown a difficult problem. In this work, we study methods to assess and bolster utterance consistency of chat systems. A dataset is first developed for studying the inconsistencies, where inconsistent dialogue responses, explanations of the inconsistencies, and recovery utterances are authored by annotators. This covers the life span of inconsistencies, namely introduction, understanding, and resolution. Building on this, we introduce a set of tasks centered on dialogue consistency, specifically focused on its detection and resolution. Our experimental findings indicate that our dataset significantly helps the progress in identifying and resolving conversational inconsistencies, and current popular large language models like ChatGPT which are good at resolving inconsistencies however still struggle with detection.

This paper introduces the Decomposed Requirements Following Ratio (DRFR), a new metric for evaluating Large Language Models’ (LLMs) ability to follow instructions. Addressing a gap in current methodologies, DRFR breaks down complex instructions into simpler criteria, facilitating a detailed analysis of LLMs’ compliance with various aspects of tasks. Alongside this metric, we present InFoBench, a benchmark comprising 500 diverse instructions and 2,250 decomposed questions across multiple constraint categories. Our experiments compare DRFR with traditional scoring methods and explore annotation sources, including human experts, crowd-sourced workers, and GPT-4. The findings demonstrate DRFR’s higher reliability and the effectiveness of using GPT-4 as a cost-efficient annotator. The evaluation of several advanced LLMs using this framework reveals their strengths and areas needing improvement, particularly in complex instruction-following. This study contributes a novel metric and benchmark, offering insights for future LLM development and evaluation.

In the past year, MultiModal Large Language Models (MM-LLMs) have undergone substantial advancements, augmenting off-the-shelf LLMs to support MM inputs or outputs via cost-effective training strategies. The resulting models not only preserve the inherent reasoning and decision-making capabilities of LLMs but also empower a diverse range of MM tasks. In this paper, we provide a comprehensive survey aimed at facilitating further research of MM-LLMs. Initially, we outline general design formulations for model architecture and training pipeline. Subsequently, we introduce a taxonomy encompassing 126 MM-LLMs, each characterized by its specific formulations. Furthermore, we review the performance of selected MM-LLMs on mainstream benchmarks and summarize key training recipes to enhance the potency of MM-LLMs. Finally, we explore promising directions for MM-LLMs while concurrently maintaining a real-time tracking website for the latest developments in the field. We hope that this survey contributes to the ongoing advancement of the MM-LLMs domain.

For a LLM to be trustworthy, its confidence level should be well-calibrated with its actual performance. While it is now common sense that LLM performances are greatly impacted by prompts, the confidence calibration in prompting LLMs has yet to be thoroughly explored.In this paper, we explore how different prompting strategies influence LLM confidence calibration and how it could be improved. We conduct extensive experiments on six prompting methods in the question-answering context and we observe that, while these methods help improve the expected LLM calibration, they also trigger LLMs to be over-confident when responding to some instances.Inspired by human cognition, we propose Fact-and-Reflection (FaR) prompting, which improves the LLM calibration in two steps. First, FaR elicits the known “facts” that are relevant to the input prompt from the LLM. And then it asks the model to “reflect” over them to generate the final answer.Experiments show that FaR prompting achieves significantly better calibration; it lowers the Expected Calibration Error by 23.5% on our multi-purpose QA tasks. Notably, FaR prompting even elicits the capability of verbally expressing concerns in less confident scenarios, which helps trigger retrieval augmentation for solving these harder instances.

This work studies mitigating fact-conflicting hallucinations for large language model (LLM) at inference time.Particularly, we propose a self-endorsement framework that leverages the fine-grained fact-level comparisons across multiple sampled responses.Compared with prior ensemble methods (e.g., self-consistency) that perform response-level selection, our approach can better alleviate hallucinations for knowledge-intensive tasks.Our approach can broadly benefit smaller and open-source LLMs as it mainly conducts simple content-based comparisons.Experiments on Biographies show that our method can effectively improve the factuality of generations with simple and intuitive prompts across different scales of LLMs.Besides, comprehensive analyses on TriviaQA and GSM8K demonstrate the potential of self-endorsement for broader application.

Dense retrieval has become a prominent method to obtain relevant context or world knowledge in open-domain NLP tasks. When we use a learned dense retriever on a retrieval corpus at inference time, an often-overlooked design choice is the retrieval unit in which the corpus is indexed, e.g. document, passage, or sentence. We discover that the retrieval unit choice significantly impacts the performance of both retrieval and downstream tasks. Distinct from the typical approach of using passages or sentences, we introduce a novel retrieval unit, proposition, for dense retrieval. Propositions are defined as atomic expressions within text, each encapsulating a distinct factoid and presented in a concise, self-contained natural language format. We conduct an empirical comparison of different retrieval granularity. Our experiments reveal that indexing a corpus by fine-grained units such as propositions significantly outperforms passage-level units in retrieval tasks. Moreover, constructing prompts with fine-grained retrieved units for retrieval-augmented language models improves the performance of downstream QA tasks given a specific computation budget.

Supervised fine-tuning enhances the problem-solving abilities of language models across various mathematical reasoning tasks. To maximize such benefits, existing research focuses on broadening the training set with various data augmentation techniques, which is effective for standard single-round question-answering settings. Our work introduces a novel technique aimed at cultivating a deeper understanding of the training problems at hand, enhancing performance not only in standard settings but also in more complex scenarios that require reflective thinking. Specifically, we propose reflective augmentation, a method that embeds problem reflection into each training instance. It trains the model to consider alternative perspectives and engage with abstractions and analogies, thereby fostering a thorough comprehension through reflective reasoning. Extensive experiments validate the achievement of our aim, underscoring the unique advantages of our method and its complementary nature relative to existing augmentation techniques.

Retrieval-augmented language model (RALM) represents a significant advancement in mitigating factual hallucination by leveraging external knowledge sources. However, the reliability of the retrieved information is not always guaranteed, and the retrieval of irrelevant data can mislead the response generation. Moreover, standard RALMs frequently neglect their intrinsic knowledge due to the interference from retrieved information. In instances where the retrieved information is irrelevant, RALMs should ideally utilize their intrinsic knowledge or, in the absence of both intrinsic and retrieved knowledge, opt to respond with “unknown” to avoid hallucination. In this paper, we introduces Chain-of-Note (CoN), a novel approach to improve robustness of RALMs in facing noisy, irrelevant documents and in handling unknown scenarios. The core idea of CoN is to generate sequential reading notes for each retrieved document, enabling a thorough evaluation of their relevance to the given question and integrating this information to formulate the final answer. Our experimental results show that GPT-4, when equipped with CoN, outperforms the Chain-of-Thought approach. Besides, we utilized GPT-4 to create 10K CoN data, subsequently trained on smaller models like OPT and LLaMa-2. Our experiments across four open-domain QA benchmarks show that fine-tuned RALMs equipped with CoN significantly outperform standard fine-tuned RALMs.

Reasoning is most powerful when an LLM accurately aggregates relevant information. We examine the critical role of information aggregation in reasoning by requiring the LLM to analyze sports narratives. To succeed at this task, an LLM must infer points from actions, identify related entities, attribute points accurately to players and teams, and compile key statistics to draw conclusions. We conduct comprehensive experiments with real NBA basketball data and present SportsGen, a new method to synthesize game narratives. By synthesizing data, we can rigorously evaluate LLMs’ reasoning capabilities under complex scenarios with varying narrative lengths and density of information. Our findings show that most models, including GPT-4o, often fail to accurately aggregate basketball scores due to frequent scoring patterns. Open-source models like Llama-3 further suffer from significant score hallucinations. Finally, the effectiveness of reasoning is influenced by narrative complexity, information density, and domain-specific terms, highlighting the challenges in analytical reasoning tasks.

Speaker identification (SI) in texts aims to identify the speaker(s) for each utterance in texts. Previous studies divide SI into several sub-tasks (e.g., quote extraction, named entity recognition, gender identification, and coreference resolution). However, we are still far from solving these sub-tasks, making SI systems that rely on them seriously suffer from error propagation. End-to-end SI systems, on the other hand, are not limited by individual modules, but suffer from insufficient training data from the existing small-scale datasets. To make large end-to-end models possible, we design a new annotation guideline that regards SI as span extraction from the local context, and we annotate by far the largest SI dataset for Chinese named CSI based on eighteen novels. Viewing SI as a span selection task also introduces the possibility of applying existing storng extractive machine reading comprehension (MRC) baselines. Surprisingly, simply using such a baseline without human-annotated character names and carefully designed rules, we can already achieve performance comparable or better than those of previous state-of-the-art SI methods on all public SI datasets for Chinese. Furthermore, we show that our dataset can serve as additional training data for existing benchmarks, which leads to further gains (up to 6.5% in accuracy). Finally, using CSI as a clean source, we design an effective self-training paradigm to continuously leverage hundreds of unlabeled novels.

We consider the problem of pretraining a two-stage open-domain question answering (QA) system (retriever + reader) with strong transfer capabilities. The key challenge is how to construct a large amount of high-quality question-answer-context triplets without task-specific annotations. Specifically, the triplets should align well with downstream tasks by: (i) covering a wide range of domains (for open-domain applications), (ii) linking a question to its semantically relevant context with supporting evidence (for training the retriever), and (iii) identifying the correct answer in the context (for training the reader). Previous pretraining approaches generally fall short of one or more of these requirements. In this work, we automatically construct a large-scale corpus that meets all three criteria by consulting millions of references cited within Wikipedia. The well-aligned pretraining signals benefit both the retriever and the reader significantly. Our pretrained retriever leads to 2%-10% absolute gains in top-20 accuracy. And with our pretrained reader, the entire system improves by up to 4% in exact match.

Comprehending a dialogue requires a model to capture diverse kinds of key information in the utterances, which are either scattered around or implicitly implied in different turns of conversations. Therefore, dialogue comprehension requires diverse capabilities such as paraphrasing, summarizing, and commonsense reasoning. Towards the objective of pre-training a zero-shot dialogue comprehension model, we develop a novel narrative-guided pre-training strategy that learns by narrating the key information from a dialogue input. However, the dialogue-narrative parallel corpus for such a pre-training strategy is currently unavailable. For this reason, we first construct a dialogue-narrative parallel corpus by automatically aligning movie subtitles and their synopses. We then pre-train a BART model on the data and evaluate its performance on four dialogue-based tasks that require comprehension. Experimental results show that our model not only achieves superior zero-shot performance but also exhibits stronger fine-grained dialogue comprehension capabilities. The data and code are available at https://github.com/zhaochaocs/Diana.

To perform well on a machine reading comprehension (MRC) task, machine readers usually require commonsense knowledge that is not explicitly mentioned in the given documents. This paper aims to extract a new kind of structured knowledge from scripts and use it to improve MRC. We focus on scripts as they contain rich verbal and nonverbal messages, and two relevant messages originally conveyed by different modalities during a short time period may serve as arguments of a piece of commonsense knowledge as they function together in daily communications. To save human efforts to name relations, we propose to represent relations implicitly by situating such an argument pair in a context and call it contextualized knowledge. To use the extracted knowledge to improve MRC, we compare several fine-tuning strategies to use the weakly-labeled MRC data constructed based on contextualized knowledge and further design a teacher-student paradigm with multiple teachers to facilitate the transfer of knowledge in weakly-labeled MRC data. Experimental results show that our paradigm outperforms other methods that use weakly-labeled data and improves a state-of-the-art baseline by 4.3% in accuracy on a Chinese multiple-choice MRC dataset C³, wherein most of the questions require unstated prior knowledge. We also seek to transfer the knowledge to other tasks by simply adapting the resulting student reader, yielding a 2.9% improvement in F1 on a relation extraction dataset DialogRE, demonstrating the potential usefulness of the knowledge for non-MRC tasks that require document comprehension.

Podcasts have shown a recent rise in popularity. Summarization of podcasts is of practical benefit to both content providers and consumers. It helps people quickly decide whether they will listen to a podcast and/or reduces the cognitive load of content providers to write summaries. Nevertheless, podcast summarization faces significant challenges including factual inconsistencies of summaries with respect to the inputs. The problem is exacerbated by speech disfluencies and recognition errors in transcripts of spoken language. In this paper, we explore a novel abstractive summarization method to alleviate these issues. Our approach learns to produce an abstractive summary while grounding summary segments in specific regions of the transcript to allow for full inspection of summary details. We conduct a series of analyses of the proposed approach on a large podcast dataset and show that the approach can achieve promising results. Grounded summaries bring clear benefits in locating the summary and transcript segments that contain inconsistent information, and hence improve summarization quality in terms of automatic and human evaluation.

Coreference resolution over semantic graphs like AMRs aims to group the graph nodes that represent the same entity. This is a crucial step for making document-level formal semantic representations. With annotated data on AMR coreference resolution, deep learning approaches have recently shown great potential for this task, yet they are usually data hunger and annotations are costly. We propose a general pretraining method using variational graph autoencoder (VGAE) for AMR coreference resolution, which can leverage any general AMR corpus and even automatically parsed AMR data. Experiments on benchmarks show that the pretraining approach achieves performance gains of up to 6% absolute F1 points. Moreover, our model significantly improves on the previous state-of-the-art model by up to 11% F1.