Papers for

robotics system engineers

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

MemFolder improves agent memory efficiency by folding records and context

MemFolder: Structure-Guided Memory Folding for Efficient Agent Memory Management

Abstract: Long-term agents accumulate large external memory stores as interactions unfold, increasing the costs of storage, retrieval, and response generation. Efficiency is constrained by two coupled factors: a growing record count increases retrieval overhead, while longer retrieved contexts increase prefill computation and decoding-time attention costs. Existing methods prune or consolidate records, but pruning can discard useful evidence, whereas consolidation may produce fewer yet longer retrieval units. We introduce MemFolder, a structure-guided memory folding framework that sequentially reduces record count and context length. The framework comprises two core components that share memory groups, using cross-record structure to guide both evidence consolidation and compact representation. RecordFolder grounds entity mentions, progressively selects representative yet complementary anchors, and combines semantic and entity affinity with temporal chaining to consolidate compatible records into fewer retrieval units. ContextFolder encodes the original records within these groups into compact text and uses verification feedback to revise encodings when key facts cannot be recovered. Experiments on the three text-based memory benchmarks LoCoMo, PersonaMem, and LongMemEval and the two multimodal benchmarks M3-Bench-web and M3-Bench-robot demonstrate the strong performance of our methods. Notably, on Mem0, RecordFolder retains 99.0% of the original average performance across the three text-based benchmarks with 50% fewer records, exceeding the strongest baseline by 1.9 percentage points in performance retention. Combined with ContextFolder, existing record compression methods further shorten their output records while retaining most of their performance.

Thu 19 NovComputation and LanguageComputer Vision and Pattern Recognition
The gist
Long-running AI agents keep growing memory stores as they interact, which slows down how fast they can find and use information. The authors introduce MemFolder, which smartly groups related memory records, making them fewer and shorter without losing important details. It uses structure like entities mentioned and timing to fold and compress memories in two steps, keeping key facts easy to retrieve. Their tests show MemFolder keeps most of the original performance while cutting memory size in half.
Open → 2611.26427v1

Edge device system speeds up vision language model tasks with smarter scheduling

HETEROVLM: Edge Multi-Request Vision-Language Model Serving on a Heterogeneous Versal SoC with AI Engines and Programmable Logic

Abstract: Edge devices in vehicles and robots serve concurrent requests with different deadlines to a vision-language model (VLM), each mixing a compute-bound encoder and prefill with a memory-bound decode. GPU serving targets datacenter throughput with per-iteration host scheduling: an urgent request waits for the running iteration, and small-batch decode costs much energy per token. Prior accelerators on Versal AI Engine (AIE) SoCs map one model onto a fixed AIE dataflow graph for one input stream, without decode or request scheduling. HETEROVLM keeps the AIE graph static and makes it schedulable. The AIE array runs the encoder, prefill and attention, and the programmable logic (PL) holds GEMV engines for small decode steps and a scheduler that orders short work units by class and deadline. Concurrent requests share one decode step that reads the weights once. An urgent request preempts at a work-unit boundary, and the paused request resumes from its results in DDR. On an AMD Versal VC1902, text and camera questions get their first tokens in 140 and 303 ms. HETEROVLM decodes a single request 1.3x and 4.0x faster than NVIDIA Tesla T4 and Jetson Orin Nano GPUs, with 61% and 73% of their energy per token, and uses 51% of the energy per token of an NVIDIA GeForce RTX 3090 Ti. The same bitstream serves multiple VLMs by loading their own instruction programs and weights. A demo is available at https://anonymous.4open.science/r/demo heterovlm-EFM2/.

Wed 18 NovHardware Architecture
The gist
Edge devices like those in cars and robots need to run vision-language models quickly when they get many requests with tight deadlines. The authors created a system called HETEROVLM that uses special chips to run parts of these models efficiently and schedule tasks to handle urgent requests faster. It improves speed and energy use compared to common GPUs and supports multiple models with the same hardware setup. This approach helps devices respond faster to vision and text questions while saving power.
Open → 2611.23878v1

3R-Adapter improves 3D reconstruction in low light and complex scenes

3R-Adapter: Retrieval, Rewiring, and Refinement for Efficient Adaptation of 3D Reconstruction Model

Abstract: Feed-forward 3D reconstruction models have demonstrated strong generalization under large-scale pre-training, yet they struggle in long-tail scenarios such as low-light reconstruction, joint human-scene reconstruction, and test-time adaptation. We observe relational attention structure collapse, manifested as the failure of cross-view correspondences, as a primary cause of these limitations. While adaptation offers a practical way to bridge this gap, existing parameter-efficient fine-tuning (PEFT) approaches mainly perform channel-wise feature adjustment and lack mechanisms to explicitly adapt relational structures, leaving correspondence failures unresolved. In this work, we formulate adaptation of 3D reconstruction models as a problem of correspondence recovery under domain shift and propose 3R-Adapter, a PEFT paradigm for long-tail 3D reconstruction scenarios. Our method consists of three components: Deformable Retrieval for task-adaptive relation candidate retrieval, Relational Rewiring for reconstructing cross-view relational structures, and Consistency Refinement for injecting the rewired relations into stable predictions. Extensive experiments on low-light reconstruction, joint human-scene reconstruction, and test-time adaptation demonstrate that our approach significantly improves robustness over existing PEFT methods while maintaining efficient lightweight adaptation. Our code is released at https://github.com/PKU-OV3-LAB/3R-Adapter.

Mon 9 NovComputer Vision and Pattern Recognition
The gist
3D models that reconstruct objects and scenes from images can struggle when conditions are unusual, like in low light or when people and scenes are mixed. The authors found that a common problem is the loss of relationships between different views, which causes errors in matching parts of the scene. They created a method called 3R-Adapter that fixes these relationships by retrieving helpful connections, rewiring attention between views, and refining results for clearer 3D reconstructions. Their approach makes existing 3D models work better in challenging cases without needing heavy retraining.
Open → 2611.11964v1

Evaluation system reveals real-world robot grasp success variations

GraspBench: Execution-Aware Cross-Robot Evaluation of 6-DoF Grasp Generation in MuJoCo

Abstract: Modern 6-DoF grasp generators are often evaluated using grasp scores, dataset-level metrics, or single-robot demonstrations, yet proposal quality alone does not determine whether a grasp is executable by a particular robot. We present GraspBench, a controlled MuJoCo protocol for evaluating complete grasp-generation pipelines after robot-specific inverse-kinematics filtering, motion planning, and fixed approach-close-lift execution. GraspBench evaluates six configured pipelines across four robot embodiments, five diagnostic objects, and a structured tabletop pose grid, yielding 540 scenarios per method and 3,240 method-scenario evaluations in total. Missing outputs are conservatively counted as failures, and success follows the binary lift criterion produced by each method's execution runner. GraspGen achieves the highest overall success rate (73.1%), followed by EconomicGrasp (63.5%), while robot- and object-conditioned analyses reveal substantial variation across embodiments and object geometries. GPG produces the largest native candidate sets but achieves 44.3% success, indicating that proposal volume alone does not guarantee robot-executable grasps. Cluster-aware paired intervals support the principal method differences while not estimating repeated-execution uncertainty. The results characterize complete configured pipelines on fixed scenarios rather than intrinsic model quality or stochastic repeatability, showing that grasping performance depends jointly on the generator, robot embodiment, scene, and execution protocol.

Thu 5 NovRoboticsArtificial Intelligence
The gist
Grasping objects with robots is tricky because not all proposed grasps can actually be done by a real robot. The authors designed GraspBench, a testing system using simulations to check if robotic grasp plans work on different robots and objects. Their study shows that the quality of grasp proposals alone doesn't guarantee success; how the robot and object shape affect execution matters a lot. This helps better understand which robot and grasping approaches perform well in realistic situations.
Open → 2611.07651v1

Depth image upsampling guided by high resolution color images

Image Guided Depth Upsampling using Anisotropic Total Generalized Variation

Abstract: In this work we present a novel method for the challenging problem of depth image upsampling. Modern depth cameras such as Kinect or Time of Flight cameras deliver dense, high quality depth measurements but are limited in their lateral resolution. To overcome this limitation we formulate a convex optimization problem using higher order regularization for depth image upsampling. In this optimization an anisotropic diffusion tensor, calculated from a high resolution intensity image, is used to guide the upsampling. We derive a numerical algorithm based on a primaldual formulation that is efficiently parallelized and runs at multiple frames per second. We show that this novel upsampling clearly outperforms state of the art approaches in terms of speed and accuracy on the widely used Middlebury 2007 datasets. Furthermore, we introduce novel datasets with highly accurate groundtruth, which, for the first time, enable to benchmark depth upsampling methods using real sensor data.

Mon 2 NovComputer Vision and Pattern Recognition
The gist
Depth cameras capture how far away things are but with blurry details. The authors created a new method to make these depth images sharper and more detailed by using clearer color pictures to guide the enhancement. Their approach uses a math technique to carefully fill in missing depth details without blurring edges. They also made new data to test this method using real camera measurements and found it works faster and more accurately than older methods.
Open → 2611.03399v1

Agentic robot fleets need careful workload and resource management

Serving Agentic Robot Fleets: Workload Characterization, KV-Cache Thrashing, and Admission Control

Abstract: Agentic robots combine a vision-language brain, a tool-using harness, and a vision-language-action (VLA) policy. Serving a fleet of these agents introduces two coupled resource constraints: the brain repeatedly revisits long, growing key-value (KV) cache states, while the VLA executes action chunks under tight control deadlines. We characterize 80 LIBERO-PRO episodes generated by GPT-6-Astra, RPent, and pi0.5 (1,862 brain turns and 1,582 VLA inference chunks), profile both layers on H800 graphics processing units (GPUs), and replay the resulting agent programs on a vLLM serving engine. Brain contexts reach 215k tokens; the median Astra request takes 7.78 s, while the intervening environment-step gap is only 1.0 s. Under a 0.25 s action-chunk deadline, one VLA GPU meets the 1% starvation target for one robot; under a 1.0 s deadline, its measured capacity reaches at least 32. In sustained brain replay, increasing concurrency from 20 to 32 robots reduces throughput from about 2.4 to 0.6 requests/s as repetitive prefill displaces generation. A static active-episode cap restores 2.24-2.25 requests/s and raises completed trace-program throughput from 1.30 to 5.00-5.55 episodes/min. Our measurements expose the workload-dependent shift between action deadlines and brain KV residency, and establish practical principles for capacity planning and admission in sustained agentic robot fleets.

Sun 1 NovRoboticsDistributed, Parallel, and Cluster Computing
The gist
Agentic robots use a complex setup combining language understanding and actions, but running many such robots at once is challenging because of two main resource limits. One is the large memory needed to remember past interactions (key-value cache), and the other is the tight timing needed to perform actions quickly. The authors studied how these robots perform on specific tasks using powerful GPUs and identified how workload and concurrency affect performance. They also suggest practical ways to manage how many robots run simultaneously to keep the system efficient.
Open → 2611.01404v1

Lunar robot swarms keep warm by moving like penguins at night

Emergent Cooperative Thermoregulation in Lunar Swarm Robotics

Abstract: Swarms of small robots are an attractive way to survey lunar caves and lava tubes. Missions such as NASA's CADRE will fly cooperating rovers for distributed mapping, but only in daylight. The 354 h lunar night drops the surface towards 95 K, and the radioisotope heaters that let single rovers survive it cannot be supplied at swarm scale. In this study we investigate whether a swarm can instead use its own members as thermal shielding. We train one policy, shared by all twenty robots, by reinforcement learning with an evolution-strategies estimator. The policy receives a single reward, once, at the end of the night: the mean temperature of the five coldest robots. Nothing in that reward refers to position, spacing, exposure or motion. Two behaviors nonetheless emerge. The swarm aggregates, and robots circulate between the sheltered core and the exposed rim, the same pair of behaviors with which emperor penguins survive the Antarctic winter. We measure the survival threshold, the smallest stored energy per robot at which every robot is still alive at sunrise. Across ten runs it is 155.6 +/- 1.3 pcm-h, where one pcm-h is 7.2 kJ of stored heat. That is 37% below a motionless huddle and within 13% of a physics-based lower bound, and at a budget where the motionless huddle loses six robots, every robot survives. The behavior needs no shared spatial reference and no robot-robot contact, and transfers from twenty robots to forty without retraining.

Thu 22 OctRoboticsMultiagent Systems
The gist
Lunar nights are extremely cold, making it hard for small robots to survive without heavy heaters. The authors show that a group of robots can keep warm by huddling together and taking turns moving from the cold outside to the warm center, just like emperor penguins do. This behavior emerged naturally when the robots were trained with reinforcement learning to maximize the coldest robots' temperature. Their approach helps every robot survive with less stored heat than other strategies, and it works even when the group size doubles.
Open → 2610.31868v1

Behavior cloning improves autonomous driving from bird's-eye view

Investigating Behavior Cloning from Few Demonstrations for Autonomous Driving based on Bird's-Eye View in Simulated Cities

Abstract: This paper investigates the use of Behavior Cloning (BC) for autonomous driving from a bird's-eye view (BEV) perspective in simulated urban environments. BC uses supervised learning to mimic expert driving behaviors. Previous works have applied BC in the CARLA simulator but did not fully address the challenges of traffic light compliance. Our approach enhances BC by integrating a kernel density estimator to adjust training sample weights based on action density, thereby improving the learning of rare but critical actions such as stopping at red lights and accelerating at green lights, specially in scenarios of scarce number of expert demonstrations. Using BEV inputs, which provide an abstract top-down view of the driving environment, our method simplifies the policy learning process. The trained convolutional neural network (CNN) outputs steering and acceleration actions based on these BEV inputs and additional state variables. Experimental results in the CARLA simulator demonstrate that our weighted BC method significantly improves driving performance, achieving higher route completion compared to standard BC. This weighted approach proved to be crucial in learning correct driving behaviors, particularly in test environments not encountered during training, highlighting its potential for enhancing autonomous vehicle navigation.

Wed 14 OctRoboticsMachine Learning
The gist
Getting self-driving cars to follow traffic lights correctly can be really hard, especially when there are only a few examples for the car to learn from. The authors worked on this problem by teaching a computer to copy expert drivers using overhead views of the road. They gave more attention to rare but important actions, like stopping at red lights, by weighting those examples higher during training. Their approach helped cars complete more driving routes correctly in a driving simulator, even in new, unseen cities.
Open → 2610.20203v1

Single compromised robot can stealthily disrupt multi-robot team plans

Lie Rarely, Lie Big: Stealthy Insider Attacks on LLM Robot Teams

Abstract: When a team of robots delegates planning and mutual trust to LLM agents, a single compromised robot can corrupt the shared outcome. We study this threat in a grounded task: a multi-robot survey in which measurements can be verified against the physical world, but every verification costs budget that would otherwise advance the mission. We treat the compromised robot as a stealthy adversary in the system-theoretic sense: it is limited not by an energy bound but by the team's own detectors. We then derive two bounds. First, the probability that the adversary's reports are verified is bounded below in terms of the degrees in the communication graph and the verification budget. Second, the map error caused by any stealthy adversary is bounded above by the value of a linear program over the adversary's bias distributions; its solution is an exchange rate between stealth budget and damage: below a critical verification level the worst stealthy attack tells rare, full-magnitude lies on the records least likely to be verified, and above it the better purchase is small biases hidden in the noise. In experiments where the honest robots are LLM agents, both bounds hold at the budget the attack actually spent. The experiments also show that which records an LLM robot re-checks is unbiased, but how much it re-checks is unpredictable.

Sat 3 OctRoboticsMultiagent Systems
The gist
When a group of robots work together using AI agents to plan and share trust, one bad robot can secretly mess up their shared maps and plans. The authors study how a sneaky robot might lie rarely but in big ways to avoid detection while causing damage. They provide mathematical limits on how often the lies might be checked and how much harm those lies can cause. Their experiments with AI-powered robots confirm these limits and show that while which parts get checked is fair, how intensively they are checked is unpredictable.
Open → 2610.04744v1

Omniact3d improves panoramic 3d object detection using equirectangular projection

OmniAct3D: Leveraging Foundation Geometry and Evidence-Grounded Reasoning for Panoramic 3D Detection

Abstract: Accurate 3D detection is essential for mobile embodied agents, while Vision Foundation Models (VFMs) offer transferable visual and geometric priors. Yet existing VFM-based 3D detectors rely on narrow-view monocular images or discrete perspective views, limiting coherent surround perception; equirectangular projection (ERP) instead encodes a continuous 360 scene in a single image. Direct transfer remains difficult because ERP organizes geometry and visual information differently, making object-relevant cues hard to model, localize, and preserve. We propose OmniAct3D, a framework that adapts perspective-trained VFM detectors to ERP while preserving transferable VFM priors. To resolve geometric mismatch, the ERP-Ray Geometry Adapter (ERGA-Ray) models spherical viewing rays and periodic spatial structure. To localize evidence in scene-wide context, the Visual-Action Reasoning Chain (VARC) grounds each hypothesis in relevant panoramic evidence and converts it into a structured geometric action. To recover local cues lost under fixed token budgets, the Appearance-Guided Heading Expert (AGHE) re-encodes object regions at higher resolution for heading estimation. Experiments show that OmniAct3D improves over the previous best 3D detector by 2.96 NDS points on Spheriverse and over the unadapted VFM baseline by 24.87 mAP points on PanoMMOcc. With target-specific geometry adaptation, VARC retains 95--98% of the same-configuration mAP, indicating reusable object-level 3D reasoning across sensing configurations. The source code will be made publicly available at https://github.com/FeiT-FeiTeng/OmniAct3D.

Fri 2 OctComputer Vision and Pattern RecognitionArtificial IntelligenceRobotics
The gist
Detecting objects in three dimensions is important for robots and other mobile devices. The authors developed OmniAct3D, a system that helps existing 3D detectors understand full 360-degree panoramic images better. They fixed problems caused by the way panoramic images are represented and added ways to focus on important details and reasoning about object positions. Their system improved detection accuracy on multiple test sets and keeps useful knowledge from previous detectors when switching to panoramic views.
Open → 2610.03015v1

Lang3DSeg uses point transformers for open 3D LiDAR segmentation

Lang3DSeg: Annotation-Free Open-Vocabulary 3D Segmentation with Point Transformers

Abstract: Accurate 3D semantic perception is critical for safe autonomous navigation. However, supervised LiDAR segmentation remains tied to closed taxonomies and to the cost of point-wise manual annotation. Open-vocabulary methods avoid that cost by projecting the output of 2D vision-language models onto LiDAR and distilling it into a 3D network. These methods rely almost exclusively on voxel-based sparse convolutions, and point transformers have so far been limited to indoor environments, where 3D data is dense and bounded. We present Lang3DSeg, which establishes a point transformer as the backbone for annotation-free open-vocabulary segmentation of outdoor 3D LiDAR, and is trained from scratch without geometric pre-training. This training paradigm necessitates addressing the inherent noise in 2D-to-3D label projections; specifically, naive projection often suffers from depth ambiguity, where points behind an object are erroneously assigned its semantic label. We therefore composite masks using an explicit class-priority rule and truncate each projected instance at the first gap in its depth distribution, correcting the projection error directly rather than averaging it over registered sequences. Lang3DSeg achieves 52.8% mIoU on nuScenes validation and 41.4% on SemanticKITTI, the highest among published annotation-free methods on both benchmarks. Every 3D semantic segmentation is on a single LiDAR sweep, and inference operates in real-time without running vision-language models.

Thu 1 OctComputer Vision and Pattern RecognitionRobotics
The gist
Understanding and labeling 3D scenes accurately is crucial for safe autonomous driving, but manually labeling each point in 3D scans is costly. The authors present Lang3DSeg, a system that transfers knowledge from 2D vision-language models to 3D LiDAR data without needing manual annotations. They overcome errors caused by projecting 2D labels onto 3D points by applying a depth-aware correction method. Their approach works outdoors in real-time and surpasses previous annotation-free methods on standard benchmarks.
Open → 2610.00855v1

Gui agents learn to act well with imperfect knowledge

Learning Reliable GUI Agents under Imperfect Priors

Abstract: GUI agents built on large language and vision-language models still struggle on unseen applications and complex multi-step tasks, as completing real GUI tasks depends on app-specific, temporally volatile operational knowledge that is scarce in pretraining corpora. Retrieval-augmented execution offers a natural remedy but faces two coupled bottlenecks: knowledge at scale is hard to acquire, and self-collected priors inevitably drift from the live environment due to version updates, promotions, ads, A/B tests, and personalization. We therefore argue that GUI agents should not pursue perfect knowledge but learn to act correctly under imperfect priors, and propose our framework that couples knowledge acquisition with noise-robust utilization: a structured exploration strategy traverses interactive elements, builds a UI state-transition graph, and synthesizes (task, trajectory) pairs via a VLM without human annotation; a noise-aware training strategy, grounded in a taxonomy of real GUI drift patterns, injects five types of realistic errors into self-explored trajectories to teach the agent to assess prior reliability before acting. Experiments on physical devices and online emulator benchmarks show that our method discovers more unique screens, covers more benchmark tasks, and more effectively rejects erroneous priors while leveraging correct ones, with accuracy gains that transfer across datasets.

Wed 30 SeptMachine Learning
The gist
Completing tasks on computer screens can be tricky for AI agents because each app changes over time and uses unique steps that the AI hasn’t learned before. The authors show that instead of needing perfect information, AI agents perform better when they learn to handle uncertainty and mistakes in what they know about the app. They developed a way for the AI to explore app screens, gather information automatically, and train itself to recognize when its knowledge might be wrong. Experiments show this helps the agents do more tasks accurately on different apps and devices.
Open → 2609.39547v1

Visual problem solving benchmark improves generative model planning

SolveEdit: Benchmarking Visual Problem Solving in Generative Models

Abstract: Machine intelligence is often evaluated through abstract reasoning problems, yet many real-world problems are visual, such as arranging objects, repairing layouts, or tracing routes. Solving these problems requires understanding a scene, inferring what must change to achieve a goal, and realizing that change without disturbing unrelated content. However, existing benchmarks mainly evaluate perception, generation, or explicitly specified transformations, leaving goal-driven visual problem solving underexplored. To bridge this gap, we introduce SolveEpIT, a benchmark for visual problem solving through scene transformation. Given an image and a goal, a model must infer a valid transformation from the request, the scene, or a visually expressed rule, then execute it while preserving unrelated content. SoLvEEDrr contains 2,728 cases. Atomic transition contracts specify required and protected conditions, enabling SoLvEScoRE to measure completion and unintended changes without a single reference output. The strongest evaluated model achieves only57.0% SolvEScore. We further introduce SolveEdiT-PLAN, a two-stage visual planner that instantiates the transition before generation. Under matched single-generation evaluation, it improves SoLvEScoRE by 9.1 points on average across three tested generators, including a gain from 57.0% to 71.6% for GPT-Image-2, without modifying the editor.

Mon 28 SeptComputer Vision and Pattern RecognitionArtificial Intelligence
The gist
Many real-world problems involve changing images to reach a goal, like moving objects without messing up the rest of the picture. The authors created a test called SolveEpIT to check if computer programs can do these tasks well. They also made a new two-step method that helps programs plan changes before making them, which works better than previous ways. Even the best current program only solved about half the problems correctly, showing this is a hard challenge.
Open → 2609.35504v1

Communication aware inference method cuts compute and latency in connected robotics

ComVLA: Communication-Aware Split Inference for VLA Models in 6G-Connected Robotics

Abstract: Connected robotics is an emerging 6G application where mobile robots follow natural-language instructions to manipulate physical objects. The Vision-Language-Action (VLA) models that enable this are too large to run on the robot; a common trend is to offload inference to the cloud. The wireless link, however, limits how much sensing data the edge can transmit per control step. Two recent lines address this constraint: semantic communication codecs compress sensor data but require channel-specific retraining, and VLA token pruners select tokens from image but ignore the channel. Our insight is that the dense semantic information contained in the language already indicates which visual tokens matter. We propose ComVLA, a framework that uses this language guidance to adapt the VLA token budget to the channel capacity. Transmitting 32 tokens instead of 512 on the LIBERO benchmark, ComVLA cuts inference compute by 74% and inference latency by 22% versus the original OpenVLA-OFT baseline, at a cost of 1.5 pp in average task success (95.4% vs. 96.9%), and it stays within the capacity budget under Rayleigh and Rician fading. These results demonstrate that co-designing VLA inference and wireless communication is a practical direction for 6G-connected robotics.

Mon 7 SeptRobotics
The gist
Robots that follow spoken or written instructions need large AI models to understand and act. These models are often too big to run on the robot itself and must rely on a wireless connection to a powerful server. The authors found a way to reduce the amount of data sent over the wireless link by using the language part of the instruction to guess which parts of the image are important. This reduces computing work and delays while keeping the robot's success rate high, even with imperfect wireless signals.
Open → 2609.07838v1