paperlist.ai

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

XDen-1K: A Density Field Dataset of Real-World Objects

Jingxuan Zhang, Tianqi Yu, Yatu Zhang, Jinze Wu, Kaixin Yao, Jingyang Liu, Yuyao Zhang, Jiayuan Gu, Jingyi Yu

A deep understanding of the physical world is essential for robotic manipulation and physically realistic simulation. While current methods, including VLM-based and other learning-based approaches, have shown promise in physical property inference, their evaluation is often hindered by the lack of physically grounded reference data. To address this gap, we introduce XDen-1K, the first large-scale multimodal dataset that provides physically grounded density field for real-world objects. XDen-1K comprises 1,000 real-world objects spanning 137 categories, with comprehensive data for each object, including a high-resolution, carefully curated 3D geometric model with part-level annotations and paired real-world biplanar X-ray scans. In addition, XDen-1K includes high-fidelity volumetric density field reconstructed from sparse biplanar X-ray views via a novel optimization framework. XDen-1K also provides a benchmark for density estimation and enables X-ray-conditioned volumetric segmentation. Experiments further demonstrate that the center-of-mass prior derived by XDen-1K can improve robotic manipulation performance. By providing real-world X-ray scans and physics-consistent density field, XDen-1K establishes a foundation for advancing physical property inference and embodied AI.

arXiv

Silhouette-based Gait Foundation Model

Dingqiang Ye, Chao Fan, Kartik Narayan, Bingzhe Wu, Chengwen Luo, Jianqiang Li, Vishal M. Patel

Gait patterns play a critical role in human identification and healthcare analytics, yet current progress remains constrained by small, narrowly designed models that fail to scale or generalize. Building a unified gait foundation model requires addressing two longstanding barriers: (a) Scalability. Why have gait models historically failed to follow scaling laws? (b) Generalization. Can one model serve the diverse gait tasks that have traditionally been studied in isolation? We introduce FoundationGait, the first scalable, self-supervised pretraining framework for gait understanding. Its largest version has nearly 0.13 billion parameters and is pretrained on 12 public gait datasets comprising over 2 million walking sequences. Extensive experiments demonstrate that FoundationGait, with or without fine-tuning, performs robustly across a wide spectrum of gait datasets, conditions, tasks (e.g., human identification, scoliosis screening, depression prediction, and attribute estimation), and even input modality. Notably, it achieves 48.0% zero-shot rank-1 accuracy on the challenging in-the-wild Gait3D dataset (1,000 test subjects) and 64.5% on the largest in-the-lab OU-MVLP dataset (5,000+ test subjects), setting a new milestone in robust gait recognition. Coming code and model: https://github.com/ShiqiYu/OpenGait.

arXiv

Object-aware graph matching network for cross-domain remote sensing image localization

Tao Liu, Kan Ren, Qian Chen

Cross-domain and cross-modal remote sensing image geo-localization remains challenging due to large appearance discrepancies and unstable semantic correspondence across heterogeneous sensors and platforms. Existing methods mainly rely on scene-level or global representations, which often struggle to achieve reliable alignment in complex environments, especially under severe modality gaps such as infrared-to-visible matching. To facilitate research in this setting, this study introduces IRVL328, a new infrared-visible remote sensing localization dataset designed to reflect challenging cross-modal variations. Meanwhile, this study proposes an object-aware graph matching framework that integrates object detection with dual-graph neural reasoning, where salient structural regions are represented as graph nodes and both inter-image correspondences and intra-image relations are jointly modeled; a training-only node alignment strategy is further introduced to enhance supervision without increasing inference complexity. Experimental results show that the proposed method achieves competitive performance on SUES-200 and strong performance on IRVL328 and DenseUAV, particularly in challenging cross-modal settings.

arXiv

MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal Prostate MRI Segmentation

Yovin Yahathugoda, Davide Prezzi, Patricia A. Gutierrez, Piyalitt Ittichaiwong, Vicky Goh, Sebastien Ourselin, Michela Antonelli

Active Surveillance (AS) is a treatment option for managing low and intermediate-risk prostate cancer (PCa), aiming to avoid overtreatment while monitoring disease progression through serial MRI and clinical follow-up. Accurate prostate segmentation is an important preliminary step for automating this process, enabling automated detection and diagnosis of PCa. However, existing deep-learning segmentation models are often trained on single-time-point, expertly annotated datasets, making them unsuitable for longitudinal AS analysis, where multiple time points and a scarcity of expert labels hinder effective fine-tuning. To address these challenges, we propose MambaX-Net, a novel semi-supervised, dual-scan 3D segmentation architecture that computes the segmentation for time point t by leveraging the MRI and the corresponding segmentation mask from the previous time point. We introduce two new components: (i) a Mamba-enhanced Cross-Attention Module, which integrates the Mamba block into cross-attention to efficiently capture temporal evolution and long-range spatial dependencies, and (ii) a Shape Extractor Module that encodes the previous segmentation mask into a latent anatomical representation for refined zone delineation. Moreover, we use a semi-supervised self-training strategy that leverages pseudo-labels generated from a pre-trained nnU-Net, enabling effective learning without expert annotations. MambaX-Net was evaluated on a longitudinal AS dataset, and results showed that it significantly outperforms state-of-the-art U-Net and Transformer-based models, achieving superior prostate zone segmentation even when trained on limited and noisy data.

arXiv

Decoupling High and Low Frequencies for Faithful Image Generation with Fine Details

Tejaswini Medi, Hsien-Yi Wang, Arianna Rampini, Margret Keuper

Latent generative models compress images into learned embeddings prior to synthesis, and the generation quality critically depends on how faithfully these embeddings preserve visual detail. We observe that while such embeddings are effective at reconstructing low frequency structure, they struggle to recover sharp high frequency details that are essential for perceptual realism. Conventional reconstruction objectives implicitly prioritize coarse structural information over high frequency content, which can lead to overly smoothed outputs and degraded visual quality in textured regions. Motivated by this observation, we propose DeBaT, a Decoupled frequency Band Tokenizer that explicitly separates the learning of low and high frequency band embeddings. This decoupling enables accurate reconstruction of fine details while preserving global coherence. Integrated into a latent diffusion based generative model, DeBaT allows for sharper and more realistic samples than previous latent tokenizers, confirming that the explicit decoupling of high and low frequency bands eases the preservation of visual details in learned embedding spaces.

arXiv

CKAA: Cross-subspace Knowledge Alignment and Aggregation for Robust Continual Learning

Lingfeng He, De Cheng, Zhiheng Ma, Huaijie Wang, Dingwen Zhang, Nannan Wang, Xinbo Gao

Continual Learning (CL) empowers AI models to continuously learn from sequential task streams. Recently, parameter-efficient fine-tuning (PEFT)-based CL methods have garnered increasing attention due to their superior performance. They typically allocate a unique sub-module for learning each task, with a task recognizer to select the appropriate sub-modules for testing images. However, due to the feature subspace misalignment from independently trained sub-modules, these methods tend to produce ambiguous decisions under misleading task-ids. To address this, we propose Cross-subspace Knowledge Alignment and Aggregation (CKAA), a novel framework that enhances model robustness against misleading task-ids through two key innovations: (1) Dual-level Knowledge Alignment (DKA): By aligning intra-class feature distributions across different subspaces and learning a robust global classifier through a feature simulation process, DKA enables the model to distinguish features from both correct and incorrect subspaces during training. (2) Task-Confidence-guided Mixture of Adapters (TC-MoA): A robust inference scheme that adaptively aggregates task-specific knowledge from relevant sub-modules based on task-confidence scores, avoiding overconfidence in misleading task-id predictions. Extensive experiments demonstrate that CKAA outperforms existing PEFT-based CL methods. Code is available at https://github.com/HHHLF/CKAA_TIP2026.

arXiv

FermatSyn: SAM2-Enhanced Bidirectional Mamba with Isotropic Spiral Scanning for Multi-Modal Medical Image Synthesis

Feng Yuan, Yifan Gao, Haoyue Li, Xin Gao

Multi-modal medical image synthesis is pivotal for alleviating clinical data scarcity, yet existing methods fail to reconcile global anatomical consistency with high-fidelity local detail. We propose FermatSyn, which addresses three persistent limitations: (1) SAM2-based Prior Encoder that injects domain-aware anatomical knowledge via LoRA$^{+}$ efficient fine-tuning of a frozen SAM2 Vision Transformer; (2) Hierarchical Residual Downsampling Module (HRDM) coupled with a Cross-scale Integration Network (CIN) that preserves high-frequency lesion details and adaptively fuses global--local representations; and (3) continuity constrained Fermat Spiral Scanning strategy within a Bidirectional Fermat Scan Mamba (BFS-Mamba), constructing an approximately isotropic receptive field that substantially reduces the directional bias of raster or spiral serialization. Experiments on SynthRAD2023, BraTS2019, BraTS-MEN, and BraTS-MET show FermatSyn surpasses state-of-the-art methods in PSNR, SSIM, FID, and 3D structural consistency. Downstream segmentation on synthesized images yields no significant difference from real-image training ($p{>}0.05$), confirming clinical utility. Code is available at https://github.com/gatina-yone/FermatSyn.

arXiv

ReynoldsFlow: Physics-Inspired Spatiotemporal Flow Representation for Video Understanding

Yu-Hsi Chen, Ching-Kai Lin, PingKong Huang, Chin-Tien Wu

Video understanding has largely relied on deep spatiotemporal architectures, including 3D convolutional networks and optical flow (OF) based models. While effective, these methods are often computationally expensive and depend on heuristic motion representations that are sensitive to illumination, scale, and structural changes. To address these limitations, we propose ReynoldsFlow, a physics-inspired representation grounded in the Reynolds transport theorem (RTT) and Helmholtz-Hodge decomposition (HHD). ReynoldsFlow decomposes motion into curl-free (CF) and divergence-free (DF) components, providing a principled and interpretable characterization of scene dynamics. By coupling intensity information with decomposed motion cues, it produces dynamics-aware, texture-preserving features that boost downstream tasks such as pose estimation, action recognition, and tiny object detection. Lightweight and modular, ReynoldsFlow can be readily integrated into existing architectures. Experiments across diverse benchmarks show that ReynoldsFlow consistently matches or surpasses existing approaches, offering improved generalizability and computational efficiency.

arXiv

Regressor-Guided Image Editing Shifts Emotion and Disengagement Timing in Social Media

Christoph Gebhardt, Robin Willardt, Seyedmorteza Sadat, Chih-Wei Ning, Andreas Brombach, Jie Song, Otmar Hilliges, Christian Holz

Internet overuse is a widespread phenomenon in today's digital society. Existing interventions, such as time limits or grayscaling, often rely on restrictive controls that provoke psychological reactance and are frequently circumvented. Building on prior work showing that emotional responses mediate the relationship between content consumption and online engagement, we investigate whether regulating the emotional impact of images can reduce online use in a non-coercive manner. We introduce and systematically analyze three regressor-guided image-editing approaches, spanning low-level attribute optimization, latent style-space optimization, and diffusion-based editing. While the first two modify low-level visual features (e.g., contrast, color), the diffusion-based method enables higher-level changes (e.g., adjusting clothing, facial features). A controlled image-rating study shows that only the diffusion-based approach shifts perceived emotion toward a neutral reference without reducing perceived quality. In a follow-up social media experiment, edited images were associated with earlier disengagement among users who left the feed.

arXiv

Your Turn: At Home Turning Angle Estimation for Parkinson's Disease Severity Assessment

Qiushuo Cheng, Catherine Morgan, Arindam Sikdar, Alessandro Masullo, Alan Whone, Majid Mirmehdi

People with Parkinson's Disease (PD) often experience progressively worsening gait, including changes in how they turn around, as the disease progresses. Existing clinical rating tools are not capable of capturing hour-by-hour variations of PD symptoms, as they are confined to brief assessments within clinic settings. Measuring gait turning angles continuously and passively is a component step towards using gait characteristics as sensitive indicators of disease progression in PD. This paper presents a deep learning-based approach to automatically quantify turning angles by extracting 3D skeletons from videos and calculating the rotation of hip and knee joints. We utilise state-of-the-art human pose estimation models, Fastpose and Strided Transformer, on a total of 1386 turning video clips from 24 subjects (12 people with PD and 12 healthy control volunteers), trimmed from a PD dataset of unscripted free-living videos in a home-like setting (Turn-REMAP). We also curate a turning video dataset, Turn-H3.6M, from the public Human3.6M human pose benchmark with 3D ground truth, to further validate our method. Previous gait research has primarily taken place in clinics or laboratories evaluating scripted gait outcomes, but this work focuses on free-living home settings where complexities exist, such as baggy clothing and poor lighting. Due to difficulties in obtaining accurate ground truth data in a free-living setting, we quantise the angle into the nearest bin $45^\circ$ based on the manual labelling of expert clinicians. Our method achieves a turning calculation accuracy of 41.6%, a Mean Absolute Error (MAE) of 34.7°, and a weighted precision WPrec of 68.3% for Turn-REMAP. This is the first work to explore the use of single monocular camera data to quantify turns by PD patients in a home setting.

arXiv

DifAttack++: Query-Efficient Black-Box Adversarial Attack via Hierarchical Disentangled Feature Space in Cross-Domain

Jun Liu, Jiantao Zhou, Jiandian Zeng, Jinyu Tian, Isao Echizen

This work investigates efficient score-based black-box adversarial attacks that achieve a high Attack Success Rate (ASR) and good generalization ability. We propose a novel attack framework, termed DifAttack++, which operates in a hierarchical disentangled feature space and significantly differs from existing methods that manipulate the entire feature space. Specifically, DifAttack++ firstly disentangles an image's latent representation into an Adversarial Feature (AF) and a Visual Feature (VF) using an autoencoder equipped with a carefully designed Hierarchical Decouple-Fusion (HDF) module. In this formulation, the AF primarily governs the adversarial capability of an image, while the VF largely preserves its visual appearance. To enable the feature disentanglement and image reconstruction, we jointly train two autoencoders for the clean and adversarial image domains, i.e., cross-domain, respectively, using paired clean images and their corresponding Adversarial Examples (AEs) generated by white-box attacks on available surrogate models. During the black-box attack stage, DifAttack++ iteratively optimizes the AF based on query feedback from the victim model, while keeping the VF fixed, until a successful AE is obtained. Extensive experimental results demonstrate that DifAttack++ achieves superior ASR and query efficiency compared to state-of-the-art methods, while producing AEs with comparable visual quality. Our code is available at https://github.com/csjunjun/DifAttackPlus.git.

arXiv

PACEShop: Evaluating Personalized, Actionable, Compositional, and Evidence-grounded Shopping Assistants

Weimin Lyu, Chen Luo, Guangrui Li, Yaochen Xie, Dhineshkumar Ramasubbu, Arief Koesdwiady, Wanqiu Long, Hansu Gu, Yutong Chen, Zheshen Wang, Dakuo Wang, Yi Liu

Shopping assistants are shifting from ranked product lists toward structured decision support, where systems must synthesize shopper context, product evidence, and next-step guidance into a coherent recommendation experience. This changes the unit of evaluation: a fluent response can still fail by ignoring shopper context, contradicting itself across components, or leaving defects too vague to localize. Existing personalization, grounding, and LLM-as-a-judge benchmarks cover pieces of this problem, but they do not define a joint evaluation target for structured shopping-assistant responses. We formulate this missing evaluation target as PACE: Personalized, Actionable, Compositional, and Evidence-grounded evaluation. We instantiate PACE with two artifacts: PACEShop, a benchmark dataset that makes the target measurable through 22,625 controlled records with structured personas, auditable evidence pools, GOOD/BAD labels, and gold defect family and location annotations; and PACEJudge, a training-free judging protocol that makes the target reportable through a structured output contract. Our experiments show that generic judges can recognize broad quality but fail to recover the diagnostic fields required for PACE; PACEShop makes these failures verifiable, and PACEJudge improves persona-source, cross-component, grounding, and family/location closure without retraining, showing that realistic shopping-assistant evaluation requires a task-matched output contract rather than only a stronger backbone or scalar prompt.

arXiv

CyrillicQA: The Influence of Phonetically Encoded Secret Language on LLM Performance

Erik Thureck, Leo S. Rdian

Due to the selection of their training data, large language models (LLMs) perform best on standard-language inputs from languages using the Latin alphabet with large speaker populations, while disadvantaging other language varieties. Nevertheless, they can also be a versatile tool for preserving precisely such endangered languages. But do they also possess the necessary creativity and capacity for abstraction to decode phonetically encoded language the same way humans do?

arXiv

When Failures Propagate: Causal Failure Attribution in Agentic Retrieval-Augmented Generation

Lauren Pothuru

Agentic retrieval-augmented generation (RAG) interleaves retrieval, reasoning, and answer generation across multiple hops. A retrieval error at hop 1 can surface only as a wrong answer at hop 3, while later retrieval can also repair the trajectory. This paper introduces AgenticRAG-FP, an interventional benchmark for causal failure attribution in agentic RAG. The benchmark injects a certified fault at a specified hop, re-executes the downstream trajectory, and evaluates diagnosers against the known intervention. Its central question is whether a post-hoc trace still identifies the injected hop after the suffix changes. In the completed strict dense Claude Haiku 4.5 sweep on 80 three-hop MuSiQue questions, coverage-based diagnosis is 0.91 at hop 1 and 0.00 at hops 2 and 3 (n=43,36,21 failed trajectories). A smaller content-corruption study changes an answer-bearing or bridge fact in topically intact evidence. At depth 2, where 18 failed cases remain after filtering, coverage-based diagnosis is 0.00 and a frozen-hop counterfactual probe is 0.67 in an exploratory pooled comparison. Depth-3 content estimates are descriptive only because they contain three failed cases. These results make propagation depth an explicit evaluation axis for diagnosing agentic RAG failures while distinguishing broad evidence of post-hoc signal loss from small-sample method comparisons.

arXiv

JuryProbe: An Empirical Consensus-Risk Diagnostic for Routing Reference-Free Factuality Judge Panels to Grounded Verification

Tianxin Zhou, Ruixi Lin

Panels of inexpensive LLM judges increasingly make accept-or-escalate decisions. In factuality settings, accepting a claim because several reference-free judges agree can create a hidden risk: agreement may reflect shared false-negative blind spots rather than independent evidence. We introduce JuryProbe, an empirical consensus-risk diagnostic for reference-free factuality judge panels, paired with a calibration-based routing policy. JuryProbe estimates consensus risk from a labeled calibration probe using false-negative-only (FN-only) judge correlation and false-consensus lift; when flagged high-risk, reference-free majority accepts are routed to the same judges with trusted references. On audited FEVER corruptions, reference-free panels show correlated false negatives (FN-only correlations 0.402 and 0.368; lifts 3.13x and 18.13x), while unanimous false consensus drops to zero under a trusted-reference best-case diagnostic on both minimal-pair and non-minimal-pair evidence. In flagged settings, the routed policy is by construction equivalent to grounding every reference-free majority accept (verified in 34/34 splits): improvement comes from accept-conditioned grounding, while the diagnostic determines whether to activate it. A fixed, pre-specified rule flags 8-10 of 10 splits across synthetic, benchmark-authored, and scientific families and 0 of 10 on a negative control, where standing down avoids 28% of reference acquisitions at a 0.004 increase in false accepts. False-accept reduction persists under weak BM25 retrieval at substantial coverage cost, while stale stand-down labels require periodic recalibration. JuryProbe provides no formal risk guarantee and does not establish reliable stand-down on natural panels; its supported contribution is an empirical diagnostic of high-risk panel error dependence.

arXiv

Open-Weight Masked Introspection: Measuring What Language Models Can Report About Their Own Computation

Emilio Ferrara

Are frontier models able to introspect about their internal states? Recent work suggests that under certain conditions a complex enough model can audit its own internals, call out what changed, and report back confidently about it. We tested that claim on eight open-weight models from seven families and found no such ability: asked whether their own computation had been altered, none answered better than chance. To test it we built Open-Weight Masked Introspection (OWMI), a framework that intervenes on residual-stream sites, attention heads and sparse-autoencoder features, then interrogates the model about the change against the null conditions an answer has to beat: sham runs where nothing was altered, impact-matched random perturbations, and a text-only observer that sees only the visible output. Over 78,000 measurements, no model's report discriminates a real intervention from a sham beyond chance (AUROC ~0.5007), and an equivalence test bounds the effect below 0.15 percentage points of AUROC. Surprisingly, all the information needed is in the models. A model fine-tuned to report this class of intervention reaches near-perfect recovery on held-out directions, and a linear probe recovers intervention presence from the same activations at 75% to 95.8% accuracy, sharpening to no held-out error at the last layer before the model speaks. In one model the signal surfaces in the confidence rather than the words: its yes-or-no report never varies, while the confidence attached to it separates intervention from sham at AUROC 0.647. The failure sits in the path from internal state to verbal report, so oversight that reads a model's own testimony needs validating against an internal reference. While our results show the inability of current open-weight models to introspect, the debate is not settled for future models.

arXiv

LiLiCorr: Lightweight Likelihood Correlation of Parallel Drafts for Speculative Decoding

Matan Rusanovsky, Yoav Miron, Roy Uziel, Omer Belhasin, Ran Zilberstein, Maor Ashkenazi, Michael Elad

Speculative decoding accelerates language-model inference by drafting future tokens that the target model verifies in parallel. A diffusion-style block head such as DFlash is an attractive drafter, predicting an entire block of future tokens in one forward pass. However, it is trained on per-position marginals rather than the joint block distribution, so the tokens it emits are individually plausible yet jointly incoherent. We introduce LiLiCorr, a Lightweight Likelihood-based model that Correlates the per-position marginal distributions a drafter already produces. It keeps the top-k tokens at each position as candidates and processes them jointly, producing for each an in and an out vector. A pair of adjacent candidates matches when the earlier one's out vector has high cosine similarity with the later one's in vector. These matches capture the block's joint structure without ever materializing the full joint distribution. One lightweight network pass produces all the vectors, and the pairwise scores are then computed in parallel as batched matrix operations, leaving only a cheap greedy walk sequential. We further co-train the drafter with LiLiCorr, so it learns to propose candidates that correlate into longer accepted sequences. Over the vanilla DFlash drafter, LiLiCorr raises acceptance length on every benchmark by 9 to 19%, while its scoring head accounts for about 2.8% of the per-block latency. Against DFlash and two concurrent methods that also restore coherence at draft time, LiLiCorr delivers the highest throughput in 70 of 72 settings: nine benchmarks at two target sizes under greedy and temperature-one decoding, and a throughput sweep over six concurrencies, two input lengths and three entropy tiers, with all systems equally optimized on a common serving stack. Extending LiLiCorr to inputs an order of magnitude longer than it was trained on preserves that lead.

arXiv

ProofJudge: Tool-Grounded LLM Evaluation of Formal Proof Quality in Mathlib

Shane Caldwell

Formal proofs in Lean 4 that pass the kernel's type checker can nonetheless vary widely in quality. We introduce ProofJudge, an agentic LLM-as-judge system that scores formal proof quality along five dimensions beyond correctness: library leverage, automation fit, structural clarity, statement quality, and Mathlib conventions. We evaluate ProofJudge on a novel dataset of 218 declarations drawn from distinct Mathlib PRs. The judge agent is grounded by tool access to the commit the PR is applied to, enabling it to query the library state when scoring. A judge is considered aligned with human preferences when it rates the version of the PR Mathlib accepted above the initial version that was sent back for revision. All six judge models evaluated recover the reviewers' preference well above chance, from 80.8% to 63.5%, and two open-weight judges reach roughly 70% at a tenth of the best judge's cost. We release the judge harness, evaluation dataset, and evaluation traces as open-source artifacts to support further research.

arXiv

SuTRA : Structurally-Unified Tokenization with Root Awareness

Vaibhav Rathore, Siddhant Gole, Dadhichi Telwadkar, Rooshil Bhatia, Maulik Ruparel, Siddharth Sureka, Neha Bhargava

Existing subword tokenizers optimize statistical compression but ignore morphological structure, particularly the relationship between roots and affixes. This is harmful for morphologically rich Indic languages, where basic units are complex orthographic syllables (aksharas) rather than letters. Frequency-based methods over-fragment words, arbitrarily splitting roots and affixes - a phenomenon we term Morphological Shattering. We propose SuTRA (Structurally-Unified Tokenization with Root Awareness), a morphology-aware algorithm that preserves akshara indivisibility and penalizes merges crossing morphological boundaries. We also release a new morphological segmentation dataset for Hindi, Marathi, and Gujarati. SuTRA reduces shattering, achieving peak gains of +14.7% in morphological alignment (Boundary F1) and +34% in semantic recoverability (Hindi) over BPE. These structural gains yield an average improvement of +8.08 chrF2 in machine translation.

arXiv

DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data

Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech

Current large language model development relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data. We introduce Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, that is trained from scratch and delivers highly competitive performance for English and sets a new state of the art for Danish using only permissible post-training data. Trained on a mixture of 161 datasets, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B, tested across 20 benchmarks for English, Math & Code and Danish. The model is available on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir

Keep browsing

More papers from top AI conferences

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