📚 arXiv AI/ML Paper Digest (2026-07-14)
Automatically collected on 2026-07-16 — 20 latest papers.
---
1. Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution
Fields: AI, NLP, Software Engineering, Systems Control Authors: Junjie Yin, Xinyu Feng arXiv: 2607.13034 | PDF
LLM agents increasingly automate multi-step engineering and informatics workflows, yet they rarely ask how much effort a task actually requires. They often follow a maximum-context-first strategy — re-reading files and dependencies they have already seen — turning a one-line edit into a full codebase audit. The paper formalizes minimal sufficient execution and the Agent Cognitive Redundancy Ratio (ACRR), and proposes E3 (Estimate, Execute, Expand): the agent estimates an initial operating point, executes the minimal viable path, and expands scope only when validation fails. On the MSE-Bench benchmark, E3 reduces cost by 85%, tokens by 91%, and files inspected by 92% while maintaining a 100% success rate.
---
2. The Seriality Gap in Video Diffusion Models
Fields: Machine Learning, Computer Vision Authors: Jorge Diaz Chao, Konpat Preechakul, Yuxi Liu, Yutong Bai arXiv: 2607.13031 | PDF
In controlled experiments on multi-ball hard-sphere dynamics, standard bidirectional video diffusion degrades as the causal chain lengthens, even with more denoising steps. In length-matched single-ball controls (no inter-ball interaction), the degradation largely disappears, isolating dependent event structure — not video length — as the cause. The authors identify this as a seriality gap: a mismatch between tasks requiring growing serial computation and what the denoising loop of video diffusion can provide. For deterministic video prediction, denoising steps add no serial computation beyond the backbone, indicating a structural barrier for video diffusion on serial reasoning and simulation.
---
3. TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale
Fields: Machine Learning, AI, Robotics Authors: Zhouchonghao Wu, Akshay Rangesh, Weixin Li, et al. arXiv: 2607.13028 | PDF
TerraZero is a procedural driving simulator and self-play training stack. A configurable C engine runs simulation on CPU and policy inference on GPU via a zero-copy path, sustaining 1.3 million agent steps per second on a single server-grade GPU — far exceeding existing object-level simulators. TerraZero uses logged data only as a source of real-world map geometry, populating each map with randomized rule-based road users and signal controllers, and randomizing agent dynamics, rewards, and sizes per episode. As the ego policy, TerraZero is the first fully learned policy to top the InterPlan long-tail benchmark.
---
4. PalmClaw: A Native On-Device Agent Framework for Mobile Phones
Fields: NLP, AI Authors: Hongru Cai, Yongqi Li, Ran Wei, Wenjie Li arXiv: 2607.13027 | PDF
Existing mobile agents mostly operate smartphones through the GUI (tapping, swiping, typing), forming long interface-dependent sequences without direct access to device capabilities. PalmClaw is an open-source agent framework that runs natively on the phone, managing sessions, memory, skills, tools, and the agent loop. It exposes device capabilities as device tools with explicit parameters, structured results, and clearly defined execution boundaries. Experiments show an 11.5% relative improvement in task success rate and a 94.9% reduction in completion time.
---
5. A Shortcut to Statistically Steady-State Turbulence with Flow Matching
Fields: Plasma Physics, Machine Learning Authors: Gianluca Galletti, Gerald Gutenbrunner, William Hornsby, et al. arXiv: 2607.13022 | PDF
Many nonlinear physical systems exhibit a transient phase before nonlinear interactions lead to a statistically steady state; direct numerical simulations must resolve the full transient to reach it. Based on an ergodicity assumption, the paper bypasses explicit time evolution by modeling the distribution of the saturated state directly. GyroFlow, a latent generative model, estimates gyro-kinetic turbulence steady-state statistics directly in 5D phase space, conditioned on dimensionless operating parameters, outperforming autoregressive, reduced-order, and other generative approaches with large speedups.
---
6. FlowWAM: Optical Flow as a Unified Action Representation for World Action Models
Fields: Robotics, Computer Vision Authors: Yixiang Chen, Peiyan Li, Yuan Xu, et al. arXiv: 2607.13017 | PDF
World Action Models (WAMs) leverage pretrained video generators for world modeling and action prediction, but a key challenge is representing actions in a form aligned with those generators. FlowWAM is a dual-stream diffusion framework using optical flow as a unified, video-native action representation: in policy mode it generates optical flow for action prediction, and in world-model mode it uses target optical flow sequences to guide future video generation. On RoboTwin manipulation, FlowWAM raises success rate to 92.94%, outperforming VLA and WAM baselines.
---
7. Audio-Native Speech Recognition with a Frozen Discrete-Diffusion Language Model
Fields: AI, Speech Authors: Harsha Vardhan Khurdula, Abhinav Kumar Singh, Yoeven D Khemlani, Vineet Agarwal arXiv: 2607.13013 | PDF
Instead of autoregressive token-by-token decoding, this work explores discrete diffusion language models that refine a whole transcript in parallel over few denoising steps. An audio-native interface is trained for DiffusionGemma (26B-parameter MoE model), with a frozen Whisper encoder providing acoustic features mapped into the embedding space by a lightweight projector; only ~42M parameters (0.16% of the backbone) are trained. The model achieves 6.6% WER on LibriSpeech test-clean, transcribing in roughly 8 parallel steps regardless of utterance length.
---
8. DermDepth: Toward Monocular Metric Scale 3D Reconstruction Models for Dermatology
Fields: Computer Vision Authors: Héctor Carrión, Narges Norouzi arXiv: 2607.13010 | PDF
The paper shows that dense monocular 3D reconstruction, metric-scale measurement, and surface-normal texture estimation are achievable for both dermoscopic and macro cases without extra hardware or multiple captures. DermDepth is the first single-view metric-scale 3D model for dermatology, and D-Synth is the first synthetic dermoscopic dataset with pixel-perfect 3D information. On real dermoscopic data, metric-scale error is corrected from over 16x to below 1.1x.
---
9. Dynamic Resource Allocation for Ensemble Determinization MCTS
Fields: AI Authors: Jakub Kowalski, Adam Ciężkowski, Artur Krzyżyński, Mark H. M. Winands arXiv: 2607.13007 | PDF
The paper proposes a family of enhancements to Ensemble Determinization MCTS, introducing two axes of dynamic resource allocation: dynamic determinization count (adjusting the number of determinization trees based on search behavior) and dynamic simulation allocation (non-uniformly distributing simulation budget across trees). Benchmarked on Jaipur, Lost Cities, and Splendor, specific configurations yield statistically significant strength improvements.
---
10. The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting
Fields: Machine Learning Authors: Mert Onur Cakiroglu, Mehmet Dalkilic, Hasan Kurban arXiv: 2607.13006 | PDF
Spectral predictability indices are increasingly read as answering a different question: whether adding context, a longer lookback, a retrieval plug-in, or a pretrained model will help. The paper argues these are not the same question — the value of context is a property of the operating point, not the series. Any metric built from the power spectrum is invariant under phase randomization, while retrieval and foundation models' beyond-second-order value is not. A label-free, configuration-level diagnostic — coverage deficit — is proposed, whose leading term measures the gain of beyond-spectral analog prediction over linear prediction; forecasts hold across seven benchmarks.
---
11. Watermark Forensics for Generative Models: An Information-Theoretic Perspective
Fields: Security, Machine Learning Authors: Xiaoyu Li, Zheng Gao, Xiaoyan Feng, et al. arXiv: 2607.13003 | PDF
A watermark can do more than flag machine-generated text: attribute it to the producing user, extract a hidden payload, or localize edits. These form a forensic ladder, and the paper asks what each rung costs in sample length n. For statistically undetectable schemes, attributing text to one of N users costs Θ(log N/h) tokens across all stationary ergodic sources, tight up to a (1+o(1)) factor — reportedly the first tight entropy-rate law for multi-user attribution.
---
12. X-Lens: Real-Time Metric Depth Estimation with Heterogeneous Cameras
Fields: Computer Vision Authors: Heng Zhou, Shuhong Liu, Yonghao He, et al. arXiv: 2607.12993 | PDF
X-lens is a compact feed-forward model for metric depth estimation from a variable number of calibrated fisheye and pinhole views. It combines learnable calibration tokens for coarse alignment between fisheye and pinhole projection spaces, with Jacobian-parameterized distortion biases injecting local projection changes into cross-attention. Using only 0.04B parameters at up to 41 FPS, it lowers AbsRel by 25.4% on OmniScene-Full while reducing parameters by 88.9%.
---
13. Controllable Generation of Diverse Dermatological Imagery for Fair and Efficient Malignancy Classification
Fields: Computer Vision Authors: Héctor Carrión, Narges Norouzi arXiv: 2607.12987 | PDF
cgDDI (controllable diverse dermatological image generation) is a hybrid framework that (1) synthesizes realistic healthy-skin samples, (2) non-parametrically maps single-sample rare lesions to new skin tones and locations, and (3) performs efficient parametric generation from as few as 10 training samples. On the DDI benchmark, malignancy classification trained on synthetic data alone reaches 86.4% accuracy, and 90.9% state-of-the-art after fine-tuning on real data.
---
14. Win by Silence: Deletion Non-Monotonicity, Autonomous Exploitation, and Typed-State Gating in LLM Plan Evaluation
Fields: AI, Software Engineering Authors: Aleh Manchuliantsau arXiv: 2607.12986 | PDF
Plan evaluators can reward a strategic plan for becoming less explicit. The paper studies this failure in a staged expected-value scorer for LLM-generated venture routes. Proposition 1 gives the score change from deleting an interior transition while retargeting its predecessor and retaining downstream value. In a fixed cohort of 26 routes, all 57 admissible deletions match the analytic identity and threshold signs, and every route has at least one deletion that improves its score. GATE rejects score release on 26/26 silenced routes (0/26 honest pauses); after rejection, 47/54 next revisions are repaired to covering structure.
---
15. Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs
Fields: AI Authors: Sen Yang, Yuen-Hei Yeung arXiv: 2607.12985 | PDF
Aligned language models routinely misreport under non-evidential incentive pressure — agreeing with a confident user or overstating certainty even when internal belief is unchanged. Cast as a failure of internal incentive-compatibility (IC), the paper presents a method for learning and certifying counterfactual report mediators that constrain reports to a causal contract: invariant to forbidden influences (pressure, prestige, reframing) and responsive to licensed influences (genuine evidence). On the witness benchmark, a two-pass fixture simultaneously achieves resistance and update of 1.00 (Wilson 95% CI [0.99, 1.00]).
---
16. FormalAnalyticGeo: A Neural-Symbolic Based Framework for Multimodal Analytic Geometry Problem Generation
Fields: AI Authors: Ruoran Xu, Wending Gao, Qiufeng Wang arXiv: 2607.12982 | PDF
Analytic geometry remains underexplored in multimodal math reasoning, largely due to scarce annotated samples. FormalAnalyticGeo is a scalable framework for fully automated multimodal analytic geometry problem generation, built around CDL (Condition Description Language) bridging free-form problem text with exact diagram rendering via a signed distance field (SDF) engine. Four specialized LLM components — generator, formalizer, measurer, and quality validator — work sequentially. Large-scale application yields AnalyticGeo7K, a dataset of over 7K validated multimodal problems.
---
17. Ensemble Controlled-Flow Filtering for Implicit Data Assimilation
Fields: Statistical Learning, Machine Learning, Numerical Analysis, Optimization Authors: Zhuoyuan Li, Yue Zhao, Ming Li arXiv: 2607.12975 | PDF
Many observation mechanisms are many-to-one, implicit, non-smooth, or simulation-only, lacking the residual structure or likelihood guidance needed by existing ensemble filters. The paper introduces implicit data assimilation, where the analysis law is defined as an energy tilt of the forecast distribution, and proposes the Ensemble Controlled-flow Filter (EnCF) implementing this update via stochastic controlled flows. Numerically, Kalman-type filters remain preferable for smooth additive Gaussian observations, while EnCF suits non-Gaussian, many-to-one, multimodal, and implicit observation models.
---
18. The Illusion of Robustness: Aggregate Accuracy Hides Prediction Flips under Task-Irrelevant Context
Fields: NLP Authors: Yanzhe Zhang, Sanmi Koyejo, Diyi Yang arXiv: 2607.12963 | PDF
In controlled settings, state-of-the-art LLMs often appear robust to task-irrelevant context in aggregate: prepending it to benchmark questions barely affects overall accuracy. But this aggregate stability masks significant per-example instability — even semantically meaningless pseudowords made of randomly combined characters can substantially flip model predictions on a small fraction of examples. The findings reveal context-induced tail risks hidden by aggregate accuracy.
---
19. Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair in Frozen Small Code Models
Fields: Software Engineering, AI, Machine Learning Authors: Mehmet Iscan arXiv: 2607.12962 | PDF
Treating a failed program as a conjecture and an execution counterexample as an oracle-relative refutation, the paper introduces PoPE (Popperian Placebo-Controlled Evaluation) to measure whether evidence falsifying LLM-generated code can be operationally used by the same model. In the prompt channel, a form-only placebo of content ablation unlocked 12 units while the live error-pattern arm unlocked 10; in the weights channel, an error-content adapter tied the no-intervention baseline 8-8 (p=1.0).
---
20. ViCo3D: Empowering LiDAR-based Collaborative 3D Object Detection with Vision Foundation Models
Fields: Computer Vision Authors: Haojie Ren, Songrui Luo, Lingfeng Wang, et al. arXiv: 2607.12959 | PDF
Collaborative 3D perception in V2X systems typically fuses bird's-eye-view (BEV) features across agents, but BEV representations from LiDAR backbones trained from scratch are geometry-dominated and lack general semantic priors. ViCo3D adapts vision foundation models to LiDAR-based collaborative perception — projecting point clouds into the BEV plane as three-channel images so DINOv2 can extract BEV spatial visual features from LiDAR input. On DAIR-V2X, the collaborative gain is up to 1.8x over previous methods.
---
*Collected automatically on 2026-07-16.*