AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

CogArena: A Multimethod Evaluation of Cognitive Ability Structure in Large Language Models

Dengzhe Hou, Lingyu Jiang, Fangzhou Lin, Kazunori D Yamada

LLM cognitive scores are increasingly summarized as per-ability profiles whose dimensions should converge across tasks, respond selectively to matched interventions, and generalize beyond the models used to define them. We introduce CogArena, a procedurally generated 13-paradigm benchmark built around a multimethod framework for determining when cognitive-task scores warrant dimensional labels across five theory-motivated groupings. Across 55 open-weight models, nearly all paradigm correlations are positive and a common axis explains about half the variance. The within-grouping advantage is small, scoring-sensitive, and uncertain across model families. In a separately frozen, fully crossed study across 12 models from six families, targeted scaffolds show a small matched-grouping advantage, but no scaffold-specific contrast survives multiplicity correction and selectivity does not improve held-out-family prediction. The frozen confirmation criterion fails. A post-hoc alternate-wording replication produces a smaller positive estimate and again fails. Together, these results support a boundary conclusion. Theory-aligned prompting produces a small in-battery diagonal tendency, but the present evidence does not establish stable five-dimensional profiles. CogArena provides a workflow joining behavioral signatures, covariance, matched interventions, and out-of-family prediction before cognitive labels are attached to model scores.

arXiv

Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Jinghao Guo, Yifei Shen, Xinjie Zhang, Xiaoyi Zhang, Haoqing Wang, Xiao Li, Peng Zhang, Xiang An, Yin Xie, Zhening Liu, Xun Guo, Jiahao Li, Shicheng Zheng, Jinglu Wang, Zongyu Guo, Wenxuan Xie, Zihan Zheng, Yuxuan Luo, Bin Li, Yan Lu

Standard vision-language models (VLMs) suffer from Moravec's paradox: they excel at complex offline visual reasoning but struggle with simple streaming perception tasks and process them inefficiently. We present Mage-VL, an efficient codec-native streaming foundation model for real-time multimodal understanding and interaction. At its core, our custom tokenizer, Mage-ViT, replaces uniform frame sampling by selectively encoding dynamic, entropy-rich regions using motion vectors and residual energy across sparse anchor (I) and predicted (P) frames. Operating at a 16 x 16 patch level, this reduces visual token consumption by over 75% while preserving spatiotemporal context. Trained from scratch on approximately 560M unlabeled images and 100M unlabeled video frames, Mage-ViT matches or outperforms flagship encoders trained on billions of image-text pairs. We establish AI4AI data pipelines encompassing prompt-code joint optimization for multimodal captioning and AI-driven performance diagnosis to guide training recipes. Furthermore, through a bio-inspired dual-system architecture - a lightweight System 1 event gate and a causal System 2 decoder - Mage-VL enables proactive streaming perception. Extensive evaluations show that Mage-VL-4B matches Qwen3-VL-4B on static tasks while achieving strong gains in video understanding and 2D/3D spatial reasoning, with up to a 3.5x wall-clock inference speedup, and comprehensively surpasses the 15B Phi-4-reasoning-vision baseline. Beyond model artifacts, we deliver seven key empirical findings covering pre-training data efficiency, variable-resolution scaling, codec system acceleration, VideoQA SFT redundancy, motion-spatial synergy, AI4AI data pipelines, and Zero-Vision SFT for multimodal RL.

arXiv

Agent Retrieval Bench: Evaluating Repository Context Retrieval for Coding Agents

Bowen Qin, Yi Xie

Modern coding agents are usually evaluated by whether they eventually produce a correct patch, but patch generation depends on an earlier context-acquisition stage: finding the repository files needed for the task. We introduce Agent Retrieval Bench, a file-level benchmark for this upstream retrieval problem. Samples are built from real coding-workflow signals and evaluated against frozen base-commit repositories, with relevance defined by what an agent needs next rather than direct query-file semantic similarity. The benchmark covers four positive-retrieval tasks: code2test, comment2context, trace2code, and edit2ripple; a fifth subset evaluates selective retrieval using natural evidence-backed no-gold cases and counterfactual wrong-repository controls. Agent Retrieval Bench contains 427 samples across 25 repositories: 345 positive examples, 50 natural no-gold examples, and 32 counterfactual controls. The corpus includes 308 base-commit snapshots, 392,000 files, and 7.9 million chunks. We evaluate lexical retrieval, RepoMap, open-source embeddings, selective abstention, and logged agent context selection. No single retrieval family dominates: Qwen3-Embedding-4B has the best sample-weighted MRR on positive samples, Qwen3-Embedding-8B the best Recall@20, and RepoMap the best budgeted context yield at 8K tokens, with task-level winners differing substantially. Selective thresholds calibrated with counterfactual controls do not improve selective success on natural no-gold cases, revealing a calibration gap. Logged trajectories also miss every gold file on 27-35 percent of samples. A controlled seed-intervention pilot finds that retrieval-derived initial context yields higher file F1 with less post-seed exploration than random non-gold context, while oracle gold context shows substantial remaining headroom.

arXiv

Self-Distillation of Hidden Layers for Self-Supervised Representation Learning

Scott C. Lowe, Anthony Fuller, Sageev Oore, Evan Shelhamer, Graham W. Taylor

The landscape of self-supervised learning (SSL) is currently dominated by generative approaches (e.g. MAE) that reconstruct raw low-level data, and predictive approaches (e.g. I-JEPA) that predict high-level abstract embeddings. While generative methods are stable due to their reliable training targets based on ground-truth data, they are computationally inefficient for high-redundancy modalities like imagery, and their training objective does not prioritize learning high-level, conceptual features. Conversely, predictive methods often suffer from training instability due to their reliance on the non-stationary targets of final-layer self-distillation. We introduce Bootleg, a method that bridges this divide by tasking the model with predicting latent representations from multiple hidden layers of a teacher. This hierarchical objective forces the model to capture features at varying levels of abstraction simultaneously. We demonstrate Bootleg significantly outperforms comparable baselines (+10% vs. I-JEPA) on frozen probe classification of ImageNet-1K, iNaturalist-21, and VTAB, and semantic segmentation of ADE20K, Cityscapes, and COCO-Stuff.

arXiv

VEMamba: Efficient Isotropic Reconstruction of Volume Electron Microscopy with Axial-Lateral Consistent Mamba

Longmi Gao, Pan Gao

Volume Electron Microscopy (VEM) is crucial for 3D tissue imaging but often produces anisotropic data with poor axial resolution, hindering visualization and downstream analysis. Existing methods for isotropic reconstruction often suffer from neglecting abundant axial information and employing simple downsampling to simulate anisotropic data. To address these limitations, we propose VEMamba, an efficient framework for isotropic reconstruction. The core of VEMamba is a novel 3D Dependency Reordering paradigm, implemented via two key components: an Axial-Lateral Chunking Selective Scan Module (ALCSSM), which intelligently re-maps complex 3D spatial dependencies (both axial and lateral) into optimized 1D sequences for efficient Mamba-based modeling, explicitly enforcing axial-lateral consistency; and a Dynamic Weights Aggregation Module (DWAM) to adaptively aggregate these reordered sequence outputs for enhanced representational power. Furthermore, we introduce a realistic degradation simulation and then leverage Momentum Contrast (MoCo) to integrate this degradation-aware knowledge into the network for superior reconstruction. Extensive experiments on both simulated and real-world anisotropic VEM datasets demonstrate that VEMamba achieves highly competitive performance across various metrics while maintaining a lower computational footprint. The source code is available on GitHub: https://github.com/I2-Multimedia-Lab/VEMamba

arXiv

MANGO: A Global Single-Date Paired Dataset for Mangrove Segmentation

Junhyuk Heo, Beomkyu Choi, Hyunjin Shin, Darongsae Kwon

Mangroves are critical for climate-change mitigation, requiring reliable monitoring for effective conservation. While deep learning has emerged as a powerful tool for mangrove detection, its progress is hindered by the limitations of existing datasets. In particular, many resources provide only annual map products without curated single-date image-mask pairs, limited to specific regions rather than global coverage, or remain inaccessible to the public. To address these challenges, we introduce MANGO, a large-scale global dataset comprising 42,703 labeled image-mask pairs across 124 countries. To construct this dataset, we retrieve all available Sentinel-2 imagery within the year 2020 for mangrove regions and select the best single-date observations that align with the mangrove annual mask. This selection is performed using a target detection-driven approach that leverages pixel-wise coordinate references to ensure adaptive and representative image-mask pairings. We also provide a benchmark across diverse semantic segmentation architectures under a country-disjoint split, establishing a foundation for scalable and reliable global mangrove monitoring.

arXiv

Operator learning for models of tear film breakup

Qinying Chen, Arnab Roy, Tobin A. Driscoll

Tear film (TF) breakup is a key driver of understanding dry eye disease, yet estimating TF thickness and osmolarity from fluorescence (FL) imaging typically requires solving computationally expensive inverse problems. We propose an operator learning framework that replaces traditional inverse solvers with neural operators trained on simulated TF dynamics. This approach offers a scalable path toward rapid, data-driven analysis of tear film dynamics.

arXiv

GFLAN: Generative Functional Layouts

Mohamed Abouagour, Eleftherios Garyfallidis

Automated floor plan generation lies at the intersection of combinatorial search, geometric constraint satisfaction, and functional design requirements -- a confluence that has historically resisted a unified computational treatment. While recent deep learning approaches have improved the state of the art, they often struggle to capture architectural reasoning: the precedence of topological relationships over geometric instantiation, the propagation of functional constraints through adjacency networks, and the emergence of circulation patterns from local connectivity decisions. To address these fundamental challenges, this paper introduces GFLAN, a generative framework that restructures floor plan synthesis through explicit factorization into topological planning and geometric realization. Given a single exterior boundary and a front-door location, our approach departs from direct pixel-to-pixel or wall-tracing generation in favor of a principled two-stage decomposition. Stage A employs a specialized convolutional architecture with dual encoders -- separating invariant spatial context from evolving layout state -- to sequentially allocate room centroids within the building envelope via discrete probability maps over feasible placements. Stage B constructs a heterogeneous graph linking room nodes to boundary vertices, then applies a Transformer-augmented graph neural network (GNN) that jointly regresses room boundaries.

arXiv

Not All Pixels Are Equal: Pixel-wise Meta-Learning for Medical Segmentation with Noisy Labels

Chenyu Mu, Guihai Chen, Xun Yang, Erkun Yang, Cheng Deng

Medical image segmentation is crucial for clinical applications, but it is frequently disrupted by noisy annotations and ambiguous anatomical boundaries, limiting its application in real-world scenarios. Existing methods often directly adapt noisy label learning techniques designed for instance classification, overlooking the pixel-wise heterogeneity in medical segmentation with its spatially and anatomically varying difficulties. Consequently, global assumptions or simple confidence metrics fail to address these local variations, leaving boundary ambiguities unresolved. To address this issue, we propose MetaDCSeg, a robust framework that dynamically learns optimal pixel-wise weights to suppress the influence of noisy labels while preserving reliable annotations. By explicitly modeling boundary uncertainty through a Dynamic Center Distance (DCD) mechanism, our approach utilizes weighted feature distances for foreground, background, and boundary centers, directing the model's attention toward hard-to-segment pixels near ambiguous boundaries. This strategy enables more precise handling of structural boundaries, which are often overlooked by existing methods, and significantly enhances segmentation performance. Extensive experiments across four benchmark datasets with varying noise levels demonstrate that MetaDCSeg outperforms existing state-of-the-art methods.

arXiv

Recurrent Autoregressive Diffusion: Global Memory Meets Local Attention

Taiye Chen, Zihan Ding, Anjian Li, Christina Zhang, Zeqi Xiao, Yisen Wang, Chi Jin

Recent advancements in video generation has shifted from bidirectional models for short videos to autoregressive ones for ultra long video generation. Previous models, which usually use sliding window attention to restrict inference cost, lack effective memory compression and retrieval for long-term generation beyond the window size, leading to issues of forgetting and spatiotemporal inconsistencies. To enhance the retention of historical information with a fixed memory budget, we additionally incorporate temporal recurrent neural network (RNN) layers into the diffusion transformer (DiT) model. Specifically, we found that a LSTM layer after attention at each DiT layer achieves comparable performance to other state-of-the-art RNN blocks, such as Test-Time Training (TTT) and Mamba2. Moreover, existing diffusion-RNN approaches often suffer from performance degradation due to training-inference gap or the lack of overlap across windows. To address these limitations, we propose a novel Recurrent Autoregressive Diffusion (RAD) framework, which leverages recurrent blocks for memory update and retrieval and preserves local details by full attention on overlapping sliding windows, with no training and inference gap. Experiments on Memory Maze and Minecraft datasets demonstrate the superiority for long video generation by our framework with global memory and local attention.

arXiv

RobustVLA: On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations

Jianing Guo, Zhenhong Wu, Chang Tu, Yiyao Ma, Xiangqi Kong, Zhiqian Liu, Jiaming Ji, Shuning Zhang, Yuanpei Chen, Kai Chen, Qi Dou, Yaodong Yang, Xianglong Liu, Huijie Zhao, Weifeng Lv, Simin Li

In Vision-Language-Actionf(VLA) models, robustness to real-world perturbations is critical for deployment. Existing methods target simple visual disturbances, overlooking the broader multi-modal perturbations that arise in actions, instructions, environments, and observations. Here, we first evaluate the robustness of mainstream VLAs under 17 perturbations across four modalities. We find (1) actions as the most fragile modality, (2) Existing visual-robust VLA do not gain robustness in other modality, and (3) pi0 demonstrates superior robustness. To build multi-modal robust VLAs, we propose RobustVLA against perturbations in VLA inputs and outputs. For output robustness, we perform offline robust optimization against worst-case action noise that maximizes mismatch in flow matching objective. This can be seen as adversarial training, label smoothing, and outlier penalization. For input robustness, we enforce consistent actions across input variations that preserve task semantics. To account for multiple perturbations, we formulate robustness as a multi-armed bandit problem and apply an upper confidence bound algorithm to automatically identify the most harmful noise. Experiments on LIBERO demonstrate our RobustVLA delivers absolute gains over baselines of 12.6% on the pi0 backbone and 10.4% on the OpenVLA backbone across all 17 perturbations, achieving 50.6x faster inference than existing visual-robust BYOVLA that requires external LLMs, and a 10.4% gain under mixed perturbations. On the real-world FR5 robot, under four types of multimodal perturbations, RobustVLA shows strong low-data performance, outperforming pi0 by 65.6% success rate with 25 demonstrations. Even with abundant demos, our method still outperform pi0 by 30% success rate. Code and demo videos available at https://github.com/gakakulicc/RobustVLA.

arXiv

SLAM-Former: Putting SLAM into One Transformer

Yijun Yuan, Zhuoguang Chen, Kenan Li, Weibang Wang, Minghui Qin, Zhijian Fang, Weicheng Zheng, Hang Zhao

We present SLAM-Former, a neural approach that integrates full SLAM capabilities into a single transformer. Similar to traditional SLAM systems, SLAM-Former comprises both a frontend and a back-end that operate in tandem. The frontend processes sequential monocular images in real-time for incremental mapping and tracking, while the backend performs global refinement to ensure a geometrically consistent result. This alternating execution allows the frontend and back-end to mutually promote one another, enhancing overall system performance. Comprehensive experimental results demonstrate that SLAM- Former achieves superior or highly competitive performance compared to state-of-the-art dense SLAM methods.

arXiv

EchoVLM: Dynamic Mixture-of-Experts Vision-Language Model for Universal Ultrasound Intelligence

Chaoyin She, Ruifang Lu, Lida Chen, Wei Wang, Qinghua Huang

Ultrasound is the preferred early cancer screening modality due to non-ionizing radiation, cost-effectiveness, and real-time imaging, yet conventional diagnosis relies heavily on physician expertise, causing significant subjectivity and limited efficiency. Vision-Language Models (VLMs) show promise but lack ultrasound-specific knowledge and multi-organ generalization. We propose EchoVLM, the first open-source 10-billion-parameter ultrasound-tailored VLM with a Mixture-of-Experts (MoE) architecture. It is infused with knowledge across seven anatomical systems, trained on 208,941 clinical cases, 1.47 million ultrasound key-frame images, and over 100 diseases or imaging findings. Supporting clinical report generation, diagnosis prediction, and Visual Question Answering (VQA), it outperforms Qwen2-VL by 7.58 BLEU-1 and 3.45 ROUGE-1 points in report generation. This work shows substantial potential for establishing a general-purpose ultrasound VLM and lays a technical foundation for clinical translation. Source code and model weights are available at https://github.com/Asunatan/EchoVLM.

arXiv

UAV-ON: A Benchmark for Open-World Object Goal Navigation with Aerial Agents

Jianqiang Xiao, Yuexuan Sun, Yixin Shao, Boxi Gan, Rongqiang Liu, Yanjin Wu, Weili Guan, Xiang Deng

Aerial navigation is a fundamental yet underexplored capability in embodied intelligence, enabling agents to operate in large-scale, unstructured environments where traditional navigation paradigms fall short. However, most existing research follows the Vision-and-Language Navigation (VLN) paradigm, which heavily depends on sequential linguistic instructions, limiting its scalability and autonomy. To address this gap, we introduce UAV-ON, a benchmark for large-scale Object Goal Navigation (ObjectNav) by aerial agents in open-world environments, where agents operate based on high-level semantic goals without relying on detailed instructional guidance as in VLN. UAV-ON comprises 14 high-fidelity Unreal Engine environments with diverse semantic regions and complex spatial layouts, covering urban, natural, and mixed-use settings. It defines 1270 annotated target objects, each characterized by an instance-level instruction that encodes category, physical footprint, and visual descriptors, allowing grounded reasoning. These instructions serve as semantic goals, introducing realistic ambiguity and complex reasoning challenges for aerial agents. To evaluate the benchmark, we implement several baseline methods, including Aerial ObjectNav Agent (AOA), a modular policy that integrates instruction semantics with egocentric observations for long-horizon, goal-directed exploration. Empirical results show that all baselines struggle in this setting, highlighting the compounded challenges of aerial navigation and semantic goal grounding. UAV-ON aims to advance research on scalable UAV autonomy driven by semantic goal descriptions in complex real-world environments.

arXiv

Hybrid AI-Physical Modeling for Penetration Bias Correction in X-band InSAR DEMs: A Greenland Case Study

Islam Mansour, Georg Fischer, Ronny Haensch, Irena Hajnsek

Digital elevation models derived from Interferometric Synthetic Aperture Radar (InSAR) data over glacial and snow-covered regions often exhibit systematic elevation errors, commonly termed "penetration bias." We leverage existing physics-based models and propose an integrated correction framework that combines parametric physical modeling with machine learning. We evaluate the approach across three distinct training scenarios - each defined by a different set of acquisition parameters - to assess overall performance and the model's ability to generalize. Our experiments on Greenland's ice sheet using TanDEM-X data show that the proposed hybrid model corrections significantly reduce the mean and standard deviation of DEM errors compared to a purely physical modeling baseline. The hybrid framework also achieves significantly improved generalization than a pure ML approach when trained on data with limited diversity in acquisition parameters.

arXiv

Deep Delta Learning

Yifan Zhang, Yifeng Liu, Mengdi Wang, Quanquan Gu

Transformer residual streams evolve through additive updates. Although a sufficiently expressive residual block can represent content replacement, standard architectures do not parameterize reading, comparison, and replacement as an explicit residual operation. We introduce Deep Delta Learning (DDL), a structured residual update that preserves the identity path while enabling target-seeking edits to the residual state. Each layer reads the current state along a learned direction, compares the resulting readout with a learned target, and writes back a gated rank-1 correction along the same direction. Closing the gate recovers the identity map, while fully opening it exactly overwrites the selected residual readout. We instantiate DDL with both scalar and expanded residual states. The expanded formulation provides multiple persistent value channels while keeping attention and MLP computation at the original model width, thereby separating residual-state capacity from backbone compute width. Controlled LLM pretraining experiments show that DDL improves language-modeling quality and average one-shot downstream performance over additive residual baselines in the reported runs, while introducing explicit memory and throughput tradeoffs. These results suggest that depth-wise delta-rule updates provide a useful inductive bias for managing Transformer residual streams.

arXiv

Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation of LLM

Chi Zhang, Changjia Zhu, Junjie Xiong, Xiaoran Xu, Lingyao Li, Yao Liu, Zhuo Lu

Large Language Models (LLMs) have revolutionized content creation across digital platforms, offering unprecedented capabilities in natural language generation and understanding. Meanwhile, they pose risks by inadvertently producing toxic, offensive, or biased content. This dual role of LLMs, both as powerful tools for text generation and as potential sources of harmful language, presents a pressing sociotechnical challenge. In this survey, we systematically review recent studies encompassing unintentional toxicity, adversarial jailbreak attacks, and comprehensive mitigation strategies. We explore LLMs' dual role as both generators of harm and enablers of safety through detection, classification, content moderation, and prevention. We propose a unified taxonomy of LLM-related harms and defenses, analyze emerging multimodal and LLM-assisted jailbreak strategies, and assess mitigation efforts, including reinforcement learning with human feedback (RLHF), prompt engineering, and safety alignment. Our review highlights the evolving landscape of LLM safety and identifies limitations in current evaluation methodologies. Ultimately, our review outlines future research directions to guide the development of robust and ethically aligned language technologies.

arXiv

When Algorithms Meet Artists: Semantic Compression and Stake-holder Marginalisation in Public AI-Art Discourse (2013-2025)

Ariya Mukherjee-Gandhi, Oliver Muellerklein

Artists occupy a paradoxical position in generative AI. Their own work trains models that now compete with them, replicate their styles, and reshape the creative economy they inhabit. Yet whether artist concerns achieve proportional representation in the public discourse that shapes AI governance remains an open empirical question. We mapped the semantic landscape of public AI-art discourse from 2013 to 2025, drawing on 1,736 text chunks from news, podcasts, legal filings, and research, and projected 252 US-based practising artists' survey responses, captured across 70 unique frames spanning five concern dimensions, into the same space. We identify what we term semantic compression, the systematic narrowing of a diverse set of stakeholder concerns into a narrow region of public meaning-space. Compression is selective. Nearly all artist statements concentrate in just two of twenty discourse topics, while most of the remaining discourse volume sits in topics with no artist voice at all. Public discourse speaks about artists in regions where artists themselves are absent. Compression also operates inside the topics that do contain artists: dozens of distinct ownership and utility positions flow into a single topic framed as a debate about aesthetic authenticity, persisting after we control for differences in writing style between surveys and media. Public discourse amplifies broad tropes about creativity while attenuating the specific, actionable regulatory claims that artists make. Amid ongoing coalition mobilisation and litigation over AI training, these findings identify a structural pattern by which primary stakeholders are rendered peripheral in the very discourse that shapes their conditions of practice.

arXiv

Fairness Is Not Enough: Auditing Competence and Intersectional Bias in AI-powered Resume Screening

Kevin T Webster

The use of publicly available generative AI systems for resume evaluation is often justified by the assumption that these tools reduce bias relative to human judgment. However, this framing leaves a prior question unresolved: whether these systems are capable of performing the evaluative task at all. This study presents a two-part audit of eight widely used AI platforms used for resume screening. Drawing on the concept of the Illusion of Neutrality, the study examines cases in which systems appear demographically unbiased because they lack the ability to meaningfully differentiate among candidates. Experiment 1 evaluates racial and gender bias using matched fictitious resumes and finds that bias persists in context-dependent and intersectional forms. Some models penalize candidates for the presence of demographic signals, while others exhibit inconsistent patterns across roles and identities under controlled conditions. Experiment 2 evaluates task competence using resume-role mismatch and keyword-manipulation tests and finds that several models that appear relatively unbiased fail to distinguish relevant from irrelevant candidate experience in relation to the target role. In these cases, outputs appear stable or neutral not because the systems are fair, but because they fail to perform meaningful evaluation. Fairness assessments alone are therefore insufficient. The paper proposes a dual-validation framework that requires auditing for both demographic bias and evaluative competence as a minimum condition for responsible deployment.

arXiv

The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation

Tianyi Men, Zhuoran Jin, Kang Liu, Jun Zhao

Multi-turn long-horizon planning is critical for foundation model agents, yet how to fundamentally improve it remains unclear. Existing models are trained on uncontrollable and opaque Internet data, making it difficult to identify how planning ability is acquired, shaped, and integrated. To address this challenge, we introduce a unified and controlled multi-turn environment that enables precise control. It allows systematically study long-horizon planning across three stages. (1) Planning ability acquisition during pre-training. We study data format, distribution, and quality. Explicit world model construction through CoT state transition modeling yields stronger long-horizon generalization. Atomic skills alone are insufficient for compositional generalization, whereas a litte long-horizon data works. Moreover, suboptimal trajectories severely impair performance because errors amplify over long horizons. (2) Planning ability shaping via GRPO and OPD post-training. Through mutual information, we distinguish general planning patterns from task-specific planning knowledge. For planning patterns, we identify three application regions of post-training: unnecessary, effective, and unsupported. OPD has a broader effective region than GRPO under low-quality and long-horizon settings, as it provides more consistent update directions. For planning knowledge, distilling unseen procedures from a teacher with different knowledge may impair student's prior world modeling without fully establishing new knowledge. (3) Planning ability integration through MOPD post-training. We show that multi-teacher on-policy distillation (MOPD) integrates capabilities by converging to shared planning-pattern across environments. Compatible patterns enable cross-environment generalization, partially shared patterns support continual learning, while completely conflicting patterns cause severe interference.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.