HF Papers Daily Picks

[HF Papers] A paper titled "Differentiable Data Augmentation with Kornia," by Jian Shi, Edgar Riba, Dmytro Mishkin, Francesc Moreno and Anguelos Nicolaou, dated Nov 19, 2020, reviews the Kornia differentiable data augmentation (DDA) module for both spatial (2D) and volumetric (3D) tensors. According to the paper, the module leverages differentiable computer vision solutions from Kornia with the aim of integrating data augmentation pipelines and strategies into existing PyTorch components, such as autograd for differentiability and optim for optimization. The authors also provide a benchmark comparing different data augmentation frameworks and a short review of a number of approaches that use Kornia DDA. The work is a preprint and its findings have not been peer reviewed.

[HF Papers] A paper titled Investor's Last Exam introduces ILE, an equity-research benchmark designed by experienced buy-side investors and fund managers, according to the authors. It combines a versioned financial database spanning 3.43 million document versions across nine source categories plus structured market data, with validated timestamps governing retrieval and computation; a task taxonomy built around professional analyst deliverables; and an expert-led scoring framework linking criteria to sources, calculations and assumptions through financial dependency graphs. Evaluation spans 86 assignments and jointly assesses research quality, forecast accuracy and simulated investment performance. The authors report that GPT-6 Astra and Claude Opus 5.5 lead overall at 68.1 and 66.9 out of 100, yet even Astra scores only 52.5 on Forecast & Model. They conclude that even state-of-the-art models fail to translate overall research capability into predictive performance. The work is a preprint and has not been peer-reviewed.

[HF Papers] A preprint titled FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution, by Shuo Yang, Xiaoze Fan, Melissa Pan, Haocheng Xi, Zhe Wang and Shanlin Sun, presents FreeToken, an edge-native MoE serving system that treats a personal machine as a unified, elastic inference platform rather than a small GPU. According to the paper, FreeToken co-designs the full serving stack, including model layout and loading, expert residency, CPU-GPU execution, agentic state reuse and runtime memory management, and continuously maps computation and model state onto actually available resources instead of committing to a fixed offloading strategy. The authors say it supports more than 20 MoE models and real coding and tool-using agents across hardware from an 8GB laptop GPU to a single workstation GPU, scaling from a 35B model on a laptop and a 284B model on a gaming desktop to the 753B GLM-5.2 on a single workstation GPU. The system is released at flashml.ai. These are unreviewed preprint findings.

[HF Papers] Xiaomi researchers introduced Xiaomi-OCR-0, a unified 0.8B model for document parsing and OCR-centric understanding, according to the paper "Xiaomi-OCR-0 Technical Report." The team said it built an approximately 170M-sample OCR-centric corpus using an automated data engine that combines expert consensus, render-based verification and targeted synthesis. Starting from Qwen3.5-0.8B, the progressive training recipe combines Q-Mask-based text anchoring, continued pretraining and mixed-task reinforcement learning (Mix-RL). The paper reports scores of 95.24 on Real5-OmniDocBench, 96.83 on OmniDocBench v1.6 and 87.94 on Wild-OmniDocBench, plus an average of 83.2 across five OCR-oriented VQA benchmarks. Ablations indicate that, with sufficient parsing training, OCR-centric understanding supervision provides additional gains for document parsing. The results come from a preprint and have not been peer-reviewed.

[HF Papers] Hair on stylized game and avatar characters is built from hair cards, artist-placed ribbons whose geometry is the visible hair and which are animated by spring bones, according to the paper HairDeck. Existing image-based methods produce realistic hair or a single fused mesh; CHARM, described by the authors as the only prior generator of stylized hair cards, decodes a hairstyle autoregressively as a long sequence of control points. Treating a card hairstyle as an unordered set of artist-defined primitives, the authors propose HairDeck, which represents each card as a single token and uses a set diffusion transformer to denoise all cards together. From a single image it generates the cards and the spring-bone chains that drive them, yielding an animatable asset that exports to downstream formats. According to the authors, on CHARM's own test set it lowers Chamfer distance by 44% and card-orientation error by more than half while running 33 times faster, and from single images it outperforms CHARM, image-to-3D pipelines and strand-based methods. The work is a preprint and has not been peer-reviewed.

[HF Papers] A paper titled SmolDocling, by Ahmed Nassar, Andres Marafioti, Matteo Omenetti, Maksym Lysak, Nikolaos Livathinos and Christoph Auer, introduces an ultra-compact vision-language model for end-to-end document conversion. The model processes entire pages by generating DocTags, a new universal markup format that captures all page elements in full context with location. According to the authors, unlike approaches relying on large foundation models or ensemble solutions built on handcrafted pipelines of specialized models, SmolDocling offers end-to-end conversion in a 256M-parameter vision-language model, capturing content, structure and spatial location of document elements. The paper says the model shows robust performance in reproducing features such as code listings, tables, equations, charts and lists across business documents, academic papers, technical reports, patents and forms. The authors also contribute new publicly sourced datasets for charts, tables, equations and code recognition, and say experiments show SmolDocling competes with vision-language models up to 27 times larger while substantially reducing computational requirements. The model is currently available and the datasets will be publicly available soon.

[HF Papers] PaddleOCR-VL-1.6, an upgraded compact document parsing model built on PaddleOCR-VL-1.5, is presented in the paper "PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training" by Zelun Zhang and co-authors. According to the authors, PaddleOCR-VL-1.5 established a strong 0.9B baseline, but its remaining errors concentrate in under-optimized regions where model behavior is unstable, data coverage is sparse or supervision is unreliable. The paper says the new version avoids indiscriminate corpus expansion and instead introduces a region-aware data optimization framework that identifies weak regions from the previous model, applies targeted enhancement and improves supervision reliability, together with a progressive post-training recipe based on curated data selection and reinforcement learning. The authors report a new state-of-the-art score of 96.33% on OmniDocBench v1.6, strong competitiveness against top-tier VLMs and a practical post-training recipe for the PaddleOCR-VL series. The results come from a preprint and have not been peer-reviewed.

[HF Papers] A paper titled MinerU2.5 by Junbo Niu, Zheng Liu, Zhuangcheng Gu and co-authors introduces a 1.2B-parameter document parsing vision-language model that the authors say achieves state-of-the-art recognition accuracy while maintaining high computational efficiency. The approach uses a coarse-to-fine, two-stage parsing strategy that decouples global layout analysis from local content recognition. In the first stage, the model performs layout analysis on downsampled images to identify structural elements, avoiding the computational overhead of processing high-resolution inputs. In the second stage, guided by the global layout, it performs targeted content recognition on native-resolution crops from the original image to preserve fine details in dense text, complex formulas and tables. The authors say they also built a data engine to generate diverse, large-scale training corpora for pretraining and fine-tuning. According to the paper, MinerU2.5 achieves state-of-the-art performance on multiple benchmarks, surpassing general-purpose and domain-specific models across recognition tasks while maintaining significantly lower computational overhead. The work is a preprint and has not been peer-reviewed.

[HF Papers] A preprint titled "Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation" by Hermann Kumbong, Ishan S. Khare, Iddah Mlauzi, Andrew Woen, Charles Ding and E. Kelly Buchanan introduces NCM, a two-stage coarse-to-fine framework in which a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes and a transformer generates one level at a time, predicting all codes within a level in parallel. According to the authors, at matched parameter scales NCM achieves better generative perplexity than the strongest diffusion language model baseline on LM1B while running 21 times faster; on OpenWebText it is competitive with the strongest diffusion model while running 10 times faster; on XSum it matches that model's ROUGE-L while being 31 times faster; and on Sudoku it outperforms autoregressive and diffusion baselines across all difficulty levels with fewer function evaluations. The work is a preprint and has not been peer-reviewed.

[HF Papers] Frontier coding agents have sharply raised the token costs of AI-assisted software engineering, and their token efficiency is usually measured on isolated tasks rather than the continuous build-and-maintain work developers do, according to the paper TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents. Aaron J. Li and co-authors propose TokenEconomizer, an external controller that cuts environment-specific token costs over a sequence of tasks by deciding which operational micro-skills to inject, which persistently irrelevant tools to prune, and when to intervene in unproductive failure loops. The authors report that across Claude Code, Codex and GitHub Copilot CLI paired with frontier models, the method reduces average token cost by 18.4-30.9% with comparable task success, with savings holding in every repository under every tested agent-model setting and generalizing to domains such as sequential web development and security auditing, where costs fall by 6.4-21.2% depending on task structure. The work is a preprint and its findings have not been peer reviewed.

[HF Papers] A preprint titled "Work Agents: Systems, Capabilities, and Optimization" by Zishan Xu, Enjun Du, Mingqi Gao, Cheng Yang, Junhao Shen and Zirong Chen presents a work-centric, system-level perspective that views work agents as systems combining foundation models and runtime harnesses to execute tasks in stateful workspaces and deliver verifiable outcomes. According to the authors, the paper introduces two dimensions, work persistence and work openness, to characterize working conditions, and classifies digital work into five forms by task scope and organization. It identifies five core capability families required for reliable work execution and examines model optimization, harness optimization and their joint optimization, then analyzes existing evaluation methods and metrics from the perspective of work outcomes. The authors also discuss future directions including sustained work delegation, adaptation to open environments and the interaction between work execution and capability improvement. The paper is a preprint and has not been peer-reviewed.

[HF Papers] A preprint dated Sept 12, 2023 by Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng and Cody Hao Yu says high-throughput serving of large language models requires batching enough requests at once, but the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically; inefficient management wastes memory through fragmentation and redundant duplication, limiting batch size. The authors propose PagedAttention, an attention algorithm inspired by classical virtual memory and paging techniques in operating systems, and build vLLM, an LLM serving system they say achieves near-zero waste in KV cache memory and flexible sharing of KV cache within and across requests to further cut memory use. According to the authors' evaluations, vLLM improves throughput of popular LLMs by 2-4x at the same latency compared with state-of-the-art systems such as FasterTransformer and Orca, with the gain more pronounced for longer sequences, larger models and more complex decoding algorithms. The source code is publicly available on GitHub. The findings are from a preprint and have not been peer-reviewed.

[HF Papers] A preprint paper introduces OpenDevin, a platform for developing AI agents that interact with the world in ways similar to a human developer: by writing code, interacting with a command line and browsing the web. According to the paper, the platform supports implementing new agents, safe interaction with sandboxed environments for code execution, coordination between multiple agents and incorporation of evaluation benchmarks. Based on currently incorporated benchmarks, the authors evaluated agents over 15 challenging tasks, including software engineering (e.g., SWE-Bench) and web browsing (e.g., WebArena). The paper says OpenDevin is released under the permissive MIT license and is a community project spanning academia and industry with more than 1.3K contributions from over 160 contributors. The work is a preprint and has not been peer-reviewed.

[HF Papers] The paper DynaFlow: Learning Free-Form Multi-Agent Orchestration, by Ruhao Liu, Qi Li, Shizun Wang and Xinchao Wang, introduces DynaFlow, a framework that integrates free-form executable workflow synthesis into the action space of the main agent. According to the authors, the agent decides when collaboration is needed and generates task-specific programs that coordinate sub-agents through sequential or parallel execution, structured information exchange and verification, with execution results and traces informing subsequent actions. The authors said only the main agent is trained, via supervised fine-tuning on distilled trajectories followed by refinement on examples filtered for unsupported premises, biased verification and information-flow defects, while all sub-agents remain fixed. They reported that across six benchmarks spanning mathematical reasoning, scientific question answering and multi-hop retrieval, DynaFlow achieves 88.08% overall accuracy in Forced mode, outperforming the strongest evaluated baseline by 10.90 percentage points with the same Qwen3.5-27B backbone. Workflow analysis found more explicit use, verification and refinement of upstream outputs after training, the authors said. This is a preprint and the findings have not been peer-reviewed.

[HF Papers] Hy-Nomas, a video world model built on an 8B diffusion backbone, supports free-form text instructions and continuous camera control across multiple interaction turns, generating variable-length segments conditioned on current inputs and cached history, according to the paper. Its central design aligns structured video captions with action-aware causal attention, so descriptive context remains accessible across segments while action instructions are restricted to their intended temporal windows. A vision-language Prompt Enhancer translates user requests and multimodal interaction histories into structured conditions for the next segment. A multi-stage curriculum progressively introduces causal generation, text interaction, camera control and training on model-generated histories, followed by diffusion distillation. The authors say human evaluation shows Hy-Nomas outperforms existing world models, and that model weights, code and benchmarks will be released with a lightweight adaptation workflow. This is a preprint; the findings have not been peer-reviewed.

Youyang Yin, Huanhuan Liu, YY and other authors propose Unlimited OCR in a technical report, using DeepSeek OCR as the baseline and replacing all attention layers in the decoder with the proposed Reference Sliding Window Attention (R-SWA). The paper states that this design, while reducing attention computation overhead, keeps the KV cache constant throughout the decoding process, thereby alleviating the problems of rising memory usage and slowing generation under long output sequences. The authors state that, by combining the high compression rate of the DeepSeek OCR encoder with the constant KV cache design, Unlimited OCR can transcribe dozens of document pages in a single forward pass at a standard maximum length of 32K. The study also states that R-SWA is a general-purpose parsing attention mechanism and, beyond OCR, is also applicable to tasks such as speech recognition and translation. The code and model weights have been made publicly available on GitHub.

[HF Papers] Guangyi Liu, Yong Liu, Jiangning Zhang and others proposed an open-source personal agent named nanoMuse in a report, describing it as the open-source counterpart to Meta's Muse and released under the GPL-3.0 license. The paper states that Muse is an agent for a single individual, with account, device, memory and continuous conversation capabilities, but that it is closed-source, deployed on the vendor's cloud and limited to a single country. The authors argue that a personal agent should be able to operate a user's accounts and devices, remember across weeks, speak up proactively when warranted, and be accountable for its own actions; it is a piece of software rather than a model, and no open-source counterpart previously existed. nanoMuse runs one agent on every device a user owns, can operate phone screens and computers, shares the same conversation through a relay that anyone can run, routes all actions through Sentinel, keeps memory as user-readable files, and lets the user choose the model. The paper states that its scale and cost figures are estimates, and lists open memory provenance, an operating-capability evaluation suite and accompanying open models as its roadmap.

[HF Papers] Shuyuan Tu, Qi Tian, Yinming Huang and others propose Prism, a dynamic sparse attention framework for training 2K native joint video-audio generation models. The paper states that the computational cost of full attention grows quadratically with resolution, and that attention dispersed onto redundant tokens dilutes the learning signal and damages pretrained priors; existing sparse attention methods either target training-free acceleration or overlook the structural characteristic that cross-modal interaction is concentrated in sounding regions. Prism organizes the token sequence into spatiotemporal macro-regions, estimates local information structure based on video feature channel variance and the feature norm of audio-to-video cross-attention, dynamically allocates block shapes for each region, and employs a hybrid block selection strategy to determine per-query sparsity. Experiments show that Prism improves training speed by 2.5x over full attention, and its generation quality also surpasses the latter.

According to the paper "AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications," authors including Dawei Gao, Zitao Li, and Yuexiang Xie propose AgentScope 1.0, which they say targets tool-based agents interacting with environments, abstracts the foundational components required for agentic applications, and provides unified interfaces and extensible modules to facilitate developers' integration of new models and advances such as MCP. The paper states that the framework constrains agent behavior with the ReAct paradigm and, based on a systematic asynchronous design, provides agent-level infrastructure, enriches human-agent and agent-agent interaction modes, and improves execution efficiency. The framework also includes built-in agents for specific scenarios and provides an extensible evaluation module with a visual interface, a runtime sandbox, and production deployment support.

[HF Papers] According to the paper "Very Large-Scale Multi-Agent Simulation in AgentScope," Xuchen Pan, Dawei Gao, Yuexiang Xie, and other authors have developed several new features for the multi-agent platform AgentScope to support very large-scale multi-agent simulation. The authors state that existing platforms face issues in large-scale simulation such as limited scalability, low efficiency, insufficient agent diversity, and cumbersome management. To address this, the research proposes an actor-based distributed underlying mechanism and provides flexible environment support, enabling parallel execution of multiple agents, centralized workflow orchestration, and interactions between agents and between agents and the environment; it also integrates easy-to-use configurable tools and an automatic background generation pipeline to simplify the creation of diverse agent background settings, and provides a web-based interface to monitor and manage large numbers of agents deployed across devices. The authors state that comprehensive simulations have been conducted to validate the effectiveness of these enhancements, and the source code has been released on GitHub.

According to the paper "MOSS-Image Technical Report," the MOSI. AI team proposes the open-source image generation system MOSS-Image, which combines high-fidelity visual synthesis with intent-aware composition. The research claims that the system pre-trains from scratch a 5B single-stream diffusion Transformer (DiT) that supports bilingual Chinese-English long-text rendering and precise layout control; it also develops an intent-aware intelligent prompt enhancement module that transforms user requests into detailed content and layout specifications. The team further proposes the long-text rendering benchmark LLTB and the layout generation benchmark RLA. The paper states that MOSS-Image achieves the best performance among open-weight models on LLTB, with an average text coverage rate of 96.3%; it ranks second in the RLA end-to-end system comparison, and its backbone network ranks first in the generator evaluation with shared enhanced prompts. The authors state that they will release the model, training recipe, and evaluation framework.

Yu Lin, Yiming Wang, Runyuan Cai and others propose the streaming MoE inference engine Edge0, and the paper claims it can run a 35B-class MoE model on a single 24GB machine at 20tok/s, with peak active memory kept within 3GiB. The authors point out that a 35B-class model has 19.5GB of weights after 4-bit quantization; sparsity only reduces the amount of computation per token, not the bytes that must reside in memory; and naive SSD offloading does not work, because the experts for layer N+1 must be selected before the output of layer N is produced, so the reads cannot be hidden behind computation in advance. Edge0 uses a pre-routing head to predict the next layer's routing one token ahead, and uses the prediction directly as the routing, so that the prefetched expert set equals the actual routed set and no experts are dropped; it also uses an unmerged recovery LoRA trained on the student path to compensate for the quality loss caused by int4 quantization and routing substitution. The paper claims that on five public benchmarks it averages within a few percentage points of its fp16 teacher model, that the 8B tier also runs in the same framework, and that the framework, checkpoints and adapters have been open-sourced.

Yu Shi, Zongliang Fu, Shuo Chen and others propose Kronos, a unified and scalable pre-training framework for financial K-line modeling; its dedicated tokenizer discretizes continuous market information into token sequences to preserve price dynamics and trading activity patterns. The authors state that the model was pre-trained with an autoregressive objective on a multi-market corpus of more than 12 billion K-line records from 45 global exchanges. According to the paper, under a zero-shot setting, Kronos improves RankIC for price series forecasting by 93% over leading time-series foundation models and by 87% over the best non-pre-trained baseline on benchmark datasets, reduces MAE for volatility forecasting by 9%, and improves the fidelity of synthetic K-line sequence generation by 22%. The pre-trained model has been made public on GitHub.

Yijia Xiao, Edward Sun, Di Luo and others propose TradingAgents, a stock trading framework inspired by trading firms, in which LLM-driven agents play specialized roles such as fundamental analysts, sentiment analysts, technical analysts, and traders with different risk preferences. According to the paper, the framework also includes bullish and bearish researchers to assess market conditions, a risk management team to monitor exposure, and traders who make decisions by synthesizing debates and historical data. The authors say that the detailed architecture and extensive experiments show that the framework outperforms baseline models in cumulative return, Sharpe ratio, and maximum drawdown, highlighting the potential of multi-agent LLM frameworks in financial trading. The code has been released on GitHub.

Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu and others propose Harness^2, a plug-and-play recursive agent framework aimed at addressing two major difficulties faced by recursive self-improvement (RSI) in open-world scenarios: test-time tasks are novel and lack a ground-truth verifier to guide online harness editing, and the RSI process is highly task-specific and difficult to generalize. The framework establishes two levels of recursion: task-level recursion uses a contrastive "propose-probe-compose" procedure to reliably guide editing without a ground-truth verifier, and can be scaled through parallel or serial recursion; framework-level recursion extracts experience from these edits and updates the continuously improving harness, enabling the agent to better adapt to future tasks. The paper states that, on three specialized benchmarks and four model-harness configurations, a single step of recursion continuously improves the performance of GLM-5 and Gemini 3.5 Flash relative to their base harnesses, with WorkBuddy-Bench improving by up to 11.5 points, JobBench by 4.5 points, and LAB by 11.8 percentage points, and the gains further increase as the task-level recursion budget increases.