WORLDWIDE_SanGuai / Research bulletin

[ worldwide.arakne.de ]

A big world. A little reading.

A three-day research bulletin by SanGuai. Drop in, read, move on.

EST. 202672H72 HOURS ON THE BOARD

RESEARCH BULLETIN

35 papers
SUBJECT / PAPERUPVOTESARRIVED
Artificial intelligenceHugging Face Daily Papers

Back to Pixels: Lifting Latent Video Diffusion into Pixel Space

The authors present PixVid, which lifts a pretrained latent video diffusion model to generate every frame directly in pixel space. It keeps the diffusion transformer and replaces the VAE decoder with a convolutional head. The authors report that predicting clean video rather than velocity, with a larger noise shift, is key to quality, and that finetuning from Wan2.1-14B is competitive with the latent baseline within about 1.3M training clips.

Read summaryLeaves in 63 hours
5UPVOTES
Systems & hardwareHugging Face Daily Papers

FAME-QP: A GPU-Accelerated Quadratic-Programming Solver for Real-Time Model Predictive Control

The authors present FAME-QP, a GPU-accelerated solver for dense quadratic programs in linear model predictive control. It combines constraint-dependent operator representations with fused single- and two-kernel iterations replayed through CUDA graphs, automatically selecting execution configurations via device calibration and problem-specific timing. It uses a cached operator bank to choose the initial penalty and updates operators lazily when problem matrices change. The authors report numerical benchmarks and simulated humanoid and quadrotor control experiments showing lower latency than competing GPU solvers while maintaining solution accuracy and control performance.

Read summaryLeaves in 63 hours
6UPVOTES
Artificial intelligenceHugging Face Daily Papers

Kronos: A Foundation Model for the Language of Financial Markets

The authors propose Kronos, a unified pre-training framework for financial candlestick data. A specialized tokenizer discretizes continuous market information into token sequences preserving price dynamics and trade activity, and the model is pre-trained with an autoregressive objective on over 12 billion K-line records from 45 global exchanges. The authors report zero-shot gains on price forecasting, volatility prediction and synthetic data generation.

Read summaryLeaves in 63 hours
65UPVOTES
Artificial intelligenceHugging Face Daily Papers

The Value of Thinking: Learning When to Reason in Hybrid Decision Models

The authors propose zero-or-T thinking, in which a single model either answers from implicit thinking in one forward pass without emitting tokens, or performs full autoregressive reasoning when that cost is worth paying. A hybrid decision model implements this with one frozen backbone and a lightweight self-routing block that estimates the value of thinking and runs the slow mode only when that value exceeds a per-request price. Calibrated Decision Policy Optimisation trains the block from episode-level returns.

Read summaryLeaves in 63 hours
7UPVOTES
Artificial intelligenceHugging Face Daily Papers

UniSpark: Exploring Hybrid Attention for Multimodal Understanding and Generation

The authors present UniSpark-9B, a unified multimodal model built on a Mixture-of-Transformer-Experts backbone that combines full attention with GatedDeltaNet linear attention. Autoregressive text prediction and visual flow matching jointly support visual question answering, text-to-image generation, instruction-guided editing and interleaved text-image prediction. The authors report 2470.3 on MME-S and 0.83 on GenEval.

Read summaryLeaves in 63 hours
13UPVOTES
Artificial intelligenceHugging Face Daily Papers

WorldDraft: If Code Builds the World, Can Video Models Just Render It?

The paper proposes decoupling 4D world construction from visual generation, treating video models as generative renderers. It introduces WorldDraft, a benchmark of 432 paired cases across 48 executable 4D world families, and evaluates world preservation and visual realization. The authors report tension between these goals across 12 models and propose WorldDraft-Primer, a visual prompting strategy.

Read summaryLeaves in 63 hours
16UPVOTES
Artificial intelligenceHugging Face Daily Papers

nanoMuse: An Open-Source Personal Agent for Every Device You Own

The report defines a personal agent through five questions and three horizons, then analyzes Meta's Muse using the company's public record and a copy of its production prompt. The authors present nanoMuse as an open-source GPL-3.0 counterpart: one agent across a person's devices, sharing a conversation over a self-hostable relay, routing actions through a Sentinel, keeping memory as readable files, and letting users choose the model.

Read summaryLeaves in 63 hours
103UPVOTES
Artificial intelligenceHugging Face Daily Papers

Unbundling Multimodal Vision Encoding

This paper examines how multimodal language models should encode visual input. The authors note that most models use a pretrained vision encoder, while some recent recipes train it from scratch or remove it. They build a controlled framework that pretrains all models on common data and compares them at matched lifetime compute. The authors report that a visual prior is unnecessary, dedicated vision capacity matters, and that capacity need not precede multimodal fusion.

Read summaryLeaves in 63 hours
16UPVOTES
Artificial intelligenceHugging Face Daily Papers

TTB: Test-time Baking for Efficient Rendering of Decoder-only View Synthesis Models

The paper proposes Test-Time Baking (TTB), which learns to bake global scene information from cross-view attention into a lightweight MLP for fast, high-fidelity novel-view decoding. The authors say they repurpose TTT's fast-weight MLP to explicitly learn the cross-view attention mapping, and disentangle context tokens into camera-only rendering inputs and image-content mapping targets to address modality misalignment. The abstract claims up to 23x rendering speedup with quality on par with state-of-the-art methods across several benchmarks.

Read summaryLeaves in 63 hours
24UPVOTES
Artificial intelligenceHugging Face Daily Papers

Ropedia Xperience 2: A Gold Standard for Scaling Physical AI Environments and Behaviors

The authors present Ropedia Xperience 2, releasing 1,589 metric-scale 3D scans and 4,230 egocentric recordings across 74 tasks totaling 131 hours. They claim each capture can be repurposed for robot learning: scenes become physically interactive simulation environments, human motion is retargeted to whole-body robot trajectories, and further behaviors come from virtual teleoperation and automated long-horizon trajectory generation. They propose this joint capture as a gold standard for physical AI data collection.

Read summaryLeaves in 63 hours
30UPVOTES
Artificial intelligenceHugging Face Daily Papers

YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality

The authors present YuE2, which links symbolic planning to audio generation: a single AR-NAR Mixture-of-Transformers first writes a readable score with melody and harmony, expands it into semantic tokens, and renders full-song audio. The abstract reports expert preference for symbolic planning and competitive benchmark and listening results, but these claims are the authors' own and were not independently verified here.

Read summaryLeaves in 63 hours
156UPVOTES
Artificial intelligenceHugging Face Daily Papers

Open Vision Banana: A Fully Open Recipe for Generative Visual Perception

The authors present Open Vision Banana, a fully open and reproducible recipe for turning pretrained image generators into unified multi-task perception systems. They analyze design choices such as color encoding, prompt design, parameter adaptation and task balancing, noting these choices interact. Built on 4B and 9B open-weight models with public training data, the authors report zero-shot Cityscapes results above Vision Banana and introduce legend ensembling at test time.

Read summaryLeaves in 63 hours
36UPVOTES
Artificial intelligenceHugging Face Daily Papers

Evolving Infra: Agent-Adaptive LLM Serving at Scale

The paper presents Evolving Infra, an agent-adaptive LLM serving framework that adjusts compute allocation and KV cache management at runtime. The authors argue that existing prefill-decode disaggregated systems offer limited runtime flexibility and rely on coarse signals that only partly capture dynamic agent workloads, causing load imbalance and eviction of reusable KV caches. The design decouples compute allocation from KV cache management, adds an online planner for the PD split, and uses an agent-aware distributed KV manager.

Read summaryLeaves in 39 hours
12UPVOTES
Artificial intelligenceHugging Face Daily Papers

Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents

The paper examines decision models in three roles: actor, judge, and teacher. The authors compare three decision models and three general LLMs as actors and judges on ALFWorld, and use JEV as a teacher on ALFWorld and WebShop. They report that models which act poorly can still teach well, since judging actions after an episode reveals consequences that acting must anticipate.

Read summaryLeaves in 39 hours
32UPVOTES
Artificial intelligenceHugging Face Daily Papers

Factored One-sided Adam-Moment for Practical and Scalable SOAP

The paper introduces FOAM, a lightweight variant of the second-order optimizer SOAP. According to the authors, FOAM applies input-only preconditioning, drops the output-side Kronecker factor, and replaces SOAP's dense rotated second moment with a factored Adam-moment estimator built from row-by-column statistics. The abstract reports validation loss comparable to SOAP and Muon baselines in LLM pre-training from 124M to 760M parameters, with half of SOAP's optimizer state.

Read summaryLeaves in 39 hours
12UPVOTES
Artificial intelligenceHugging Face Daily Papers

The Narrative Gap: Can LLMs Help Us Navigate Diverse Narratives Across Languages?

The authors propose six capabilities for multilingual information seeking across competing narratives and introduce NarrativeBench, built from 1,333 news articles in 23 publication languages across 23 contested topics and 460 Reddit-grounded queries. They report evaluating 14 LLMs across 24 query languages and conclude that current models struggle to present or preserve diverse narratives, even when users explicitly request diversity.

Read summaryLeaves in 39 hours
2UPVOTES

Auriga: Harness-Governed LLM Agents for Effective Throughput Optimization in Execute-Order-Validate Permissioned Blockchains

The paper presents Auriga, a non-invasive framework that jointly optimizes raw throughput and transaction success rate in Execute-Order-Validate blockchains such as Hyperledger Fabric. A harness-governed multi-agent LLM architecture handles configuration search, saturation-aware evaluation, and experience consolidation, while an offline LLM extracts symbolic AST templates so a runtime scheduler can predict read-write sets and partition transactions into conflict-free groups before execution.

Read summaryLeaves in 16 hours
--UPVOTES

Caprese: Change Impact Recommendation Using History and Runtime Analysis

The abstract presents Caprese, an open-source framework for function-level change impact recommendation in Node.js. Given a change set, it ranks potentially affected functions using three modes: confidence-guided mining of recurring co-change patterns, runtime-derived dependencies such as calls, callbacks, events, timers and shared state, and a hybrid of both. The authors report an evaluation on ten open-source applications and 100 change sets, where 77.4% of relevant functions in the top 60 hybrid recommendations were found by only one analysis.

Read summaryLeaves in 16 hours
--UPVOTES

CuEVM: Scaling Smart Contract Execution on GPUs

The authors present CuEVM, a GPU-native execution engine for the Ethereum Virtual Machine that treats transaction execution as massively parallel data processing, using new data structures and GPU-aware scheduling. They report complete EVM semantics on the GPU, passing all VMTests and 96.2% of the execution-test suite, and over 8 million ERC-20 transfers per second on one commodity GPU.

Read summaryLeaves in 16 hours
--UPVOTES
Artificial intelligenceHugging Face Daily Papers

Duo: A Reciprocal World Model for Full-Duplex Interaction

This report introduces the Reciprocal World Model (RWM) and Duo v1, a full-duplex world model that continuously perceives audiovisual input and responds to users in real time. The authors treat intelligence and physical reality as equally fundamental, framing world interaction through Intent and Realization, and describe autoregressive reasoning, synchronized tokens on a unified timeline, and dual rollout distillation as key designs.

Read summaryLeaves in 16 hours
35UPVOTES

Priority-Based Network Access with Strict-Priority Scheduling in PLCA Multidrop Networks

This paper addresses Physical Layer Collision Avoidance (PLCA) in 10BASE-T1S multidrop Ethernet, arguing that PLCA's node-based cyclic access ignores traffic priority, raising latency for time-sensitive traffic and limiting scalability. The authors propose priority-based network access and scheduling that decouples medium access from node ordering, enables class-based prioritization, and can emulate Strict Priority scheduling, with analytical worst-case latency bounds.

Read summaryLeaves in 16 hours
--UPVOTES

VASTRA: Efficient Agentic Top-K Interactive Video Analysis

The authors present VASTRA, an agentic video retrieval system that lets query complexity, not an open-ended reasoning loop, drive execution: easy queries are ranked from pre-ingested summaries, while harder ones trigger targeted verification only where relevance is uncertain. The abstract reports results on 40 benchmark queries, including a fast preset that is 12.2x quicker and 10.4x cheaper than the best ReAct agent at a 5.0-point mAP reduction.

Read summaryLeaves in 16 hours
--UPVOTES

Typify: A Tool for Usage-driven Type Inference in Python

The authors present Typify, a lightweight static type inference engine for Python that requires neither existing annotations nor training data. It combines symbolic execution, fixpoint analysis and context-matching retrieval to propagate types across modules, functions and recursive calls, and ships as a command-line interface plus a VS Code extension. The authors report that it outperforms static baselines and runs about 10x faster than hybrid deep learning approaches.

Read summaryLeaves in 16 hours
--UPVOTES

Incrementally Verifiable Oracle Aided Computation

The paper addresses incrementally verifiable computation (IVC) for computations that depend on external oracles, motivated by AI tasks. Under the LWE assumption, the authors build SNARGs and extend them to IVC in two settings: robust bounded-space computations with a matching lower bound, and signed oracles yielding publicly verifiable schemes with oracle-free setup and verification.

Read summaryLeaves in 15 hours
--UPVOTES
Artificial intelligenceHugging Face Daily Papers

G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation

The paper proposes G-STAR, a spatial–group co-scaling framework for discrete autoregressive image generation. Based on product quantization, the authors argue that under a fixed nominal information budget, group size should scale quadratically with per-axis spatial compression, enabling 32× and even effective 64× compression. They also introduce a semi-autoregressive decoder combining parallel feature computation with lightweight causal modeling. The authors report competitive results on five text-to-image benchmarks and a 20× speedup, generating each 1024×1024 image in 2.5 seconds on one NVIDIA H200.

Read summaryLeaves in 15 hours
40UPVOTES
Artificial intelligenceHugging Face Daily Papers

PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models

The authors present PhysLoop, a harness that guides and constrains large language models across video generation, selection and correction, aiming to distill physical knowledge into physically consistent videos. It supports controllable training-data construction and inference-time repair. The abstract reports a release of 100K videos, generation tools and two evaluation datasets, with evaluation on the external Physics-IQ Verified benchmark.

Read summaryLeaves in 15 hours
42UPVOTES

CommJEPA: When Predictive Representation Learning Meets Wireless Intelligence

The paper introduces CommJEPA, a communication-oriented framework for predictive representation learning based on Joint-Embedding Predictive Architecture. The authors argue that much wireless AI is task-specific and reactive, while JEPA can learn latent predictive representations of channel and environmental dynamics. They also argue that wireless properties such as partial observability, multiple time scales, action-dependent evolution and resource constraints should motivate communication-aware JEPA designs.

Read summaryLeaves in 15 hours
--UPVOTES
Artificial intelligenceHugging Face Daily Papers

HiVe-Mem: Interleaving Horizontal Memory Expansion and Vertical Evidence Composition for Multimodal Agents

The paper introduces HiVe-Mem, a multimodal memory framework that interleaves horizontal memory expansion with vertical evidence composition. Horizontally, memory episodes are organized into a relational graph using textual and visual anchors, expanding semantically retrieved seeds with related neighbors. Vertically, each retrieved memory retains multiple evidence forms, and a lightweight trainable selector composes an evidence subset per memory. The authors report gains across three benchmarks and roughly 90% lower token use.

Read summaryLeaves in 15 hours
3UPVOTES

Discovering the Relation Between Architectural Models and Reference Architectures

This paper addresses checking whether concrete software architectures are consistent with their reference architecture. The authors report that general-purpose model-comparison algorithms are inaccurate for reference architectures, and propose RAMD, an LLM-based workflow combining model analysis with structured prompting and iterative compliance checking. They evaluate it on 64 open-source architectural models across IoT, self-driving, and smart parking reference architectures, claiming 85.8% mapping accuracy with GPT-4o.

Read summaryLeaves in 15 hours
--UPVOTES
Artificial intelligenceHugging Face Daily Papers

BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model

BirdDreams is an autoregressive driving world model that jointly generates perspective video and a schematic bird's-eye-view video, so scene layout becomes part of the predicted future. Two video diffusion transformers exchange information through joint attention, and the predicted BEV provides a geometric interface for ego-trajectory control. At inference it uses a single camera image without an HD map or vehicle-state inputs.

Read summaryLeaves in 15 hours
45UPVOTES

Janus: One Agent, Two Roles for Digital Twin Construction and Network Orchestration

The authors propose Janus, an agentic LLM-based framework that closes the loop between digital twin construction and network orchestration. Its NetTwin agent uses a one-agent, two-role paradigm to build a twin from partial observations while making task-driven decisions, supported by planner, modeler and coder roles plus a unified prompt module for graph reasoning.

Read summaryLeaves in 15 hours
--UPVOTES
Artificial intelligenceHugging Face Daily Papers

DeepStress: Stress-Testing Deep Search Agents

The paper proposes DeepStress, a dynamic stress-testing framework that exposes search agents to degraded evidence at controlled rates. A synthetic environment generates documents per query while independently controlling trustworthiness, relevance, and factuality. The authors add a five-way classification of final-answer outcomes and two reliability-aware metrics, RAS and CoPRAS, which credit cautious behavior and account for computational cost. Evaluating 12 systems on HotpotQA and BrowseComp-Plus questions, the authors report that several open-weight agents answer confidently despite obvious credibility red flags, and that many HotpotQA questions are answered from parametric knowledge even when all returned documents are factually wrong.

Read summaryLeaves in 15 hours
4UPVOTES

Designing Collaborative Virtual Environments for Software Engineering Education

This article presents MetaSEE, an approach linking requirements guidance, a Web XR feature model, a reusable UML component, and a browser-based multiuser platform. The authors synthesize documented formative evaluations of the framework, feature model, UML component, and platform, reporting participation counts. Participants saw potential value in shared modeling and simulation but raised navigation, interaction, visualization, and distraction concerns. A numerical audit reconstructs a student System Usability Scale mean that conflicts with the reported value.

Read summaryLeaves in 15 hours
--UPVOTES
Artificial intelligenceHugging Face Daily Papers

Structured Confidence Perturbation via Fisher Geometry for Adversarial Training

The paper argues that adversarial training suffers from pseudo-attacks, perturbations that raise loss without truly altering the model's confidence structure. The authors propose two complementary approaches: a hybrid objective combining cross-entropy and KL divergence, and a geometric reconstruction of adversarial training that maximizes distribution distance on the confidence manifold. They claim the optimal perturbation direction aligns with principal eigenvectors of the Fisher information matrix, plus an efficient algorithm avoiding high-dimensional matrix computation.

Read summaryLeaves in 15 hours
2UPVOTES
Artificial intelligenceHugging Face Daily Papers

SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models

The authors introduce SO-Arena, a benchmark covering 11 domains and five capability families to assess System One models. It uses reliable sources and manually verified reference labels, domain-specific difficulty criteria, and sandboxed interactive tasks, and also measures decision stability under controlled option-set changes and missing-answer recognition. Across 21 models, the authors report that scaling and extra reasoning help unevenly by domain, and that the top-performing model ranks only seventh in reliability.

Read summaryLeaves in 15 hours
72UPVOTES
Last collection: 16/11, 06:0172 HOURS ON THE BOARD