Multimodal Reasoning, Agentic Code Repair, and Spatio-Temporal Model Safety

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering automated reasoning and problem generation, embodied AI and robotics, safety and alignment evaluation, and domain-specific applications of learning systems.

1. FormalAnalyticGeo: A Neural-Symbolic Based Framework for Multimodal Analytic Geometry Problem Generation

Authors: Ruoran Xu, Wending Gao, Qiufeng Wang | Categories: cs.AI, cs.MA, cs.SC Link: arxiv.org/abs/2607.12982v1

This paper introduces FormalAnalyticGeo, a scalable framework for automatically generating multimodal analytic geometry problems. It uses a formal intermediate representation (CDL) and a four-component LLM pipeline—Generator, Formalizer, Measurer, and Quality Verifier—to produce 7,000+ verified problems with aligned text, diagrams, and ground truth. The system achieves a median ground-truth relative error of 0.70%.

Takeaway: A clever neural-symbolic approach that solves the data scarcity problem in analytic geometry, with a closed-loop verification system that eliminates human annotation entirely.

2. Verifier-Based Reinforcement Fine-Tuning of Reasoning Models for Thermal Energy Storage Control

Authors: Takumi Shioda, Kohei Terashima, Tatsuo Nagai | Categories: cs.LG Link: arxiv.org/abs/2607.12856v1

This study adapts an open-weight reasoning model for building thermal energy storage scheduling using reinforcement learning with verifiable rewards derived from dynamic programming. With only 30 training prompts, the model reduces emissions from 70.5 to 61.2 kg-CO2, approaching the optimal DP solution of 60.8 kg-CO2. The fine-tuned model generalizes to forecast errors and unseen conditions, though transfer to battery tasks is limited.

Takeaway: Demonstrates that RL with verifiable rewards can efficiently adapt reasoning models to specific engineering control tasks, offering a practical alternative to MPC and RL that scales across buildings.

3. Epistemic Stance Flexibility Probing: Measuring Prompt-Conditioned Register Shift in Large Language Models

Authors: Binwen Liu, Yilin Ren | Categories: cs.CL Link: arxiv.org/abs/2607.12739v1

ESFP is a behavioral benchmark that tests whether LLMs appropriately shift between neutral attribution and stance expression depending on prompt framing. Evaluating eight frontier models, the authors find epistemic flexibility is largely orthogonal to general capability: a 27B open-weight model matches proprietary systems, and reasoning-optimized models don’t consistently show higher flexibility.

Takeaway: Introduces a nuanced evaluation dimension beyond accuracy or safety—measuring whether models actually distinguish “what experts believe” from “what the model believes” in a coherent way.

4. Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference

Authors: Zebin Yang, Qi Wang, Yunhe Wang, Xiurui Guo, Bo Yu et al. | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2607.12659v1

Jetson-PI enables efficient VLA model deployment on low-power devices like the Jetson Orin, achieving 8.66x and 5.41x improvements in control frequency over naive PyTorch and vla.cpp respectively. It uses a future correction module to address perception-execution misalignment from asynchronous inference, plus confidence-based scheduling optimization and system-level accelerations.

Takeaway: Solves a critical bottleneck for real-world embodied AI—getting vision-language-action models to run fast enough on edge hardware without sacrificing task success rate.

Authors: Hongbo Wang, Huaibo Huang, Jie Cao, Jin Liu, Haoyang Tong et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2607.12752v1

Hallo4D introduces a generation-detection-correction paradigm using large multimodal models to identify and fix spatial and temporal inconsistencies in 3D and 4D content generation. The framework includes motion-aware keyframe sampling, LMM-guided initialization, and consensus-driven optimization via multi-model voting, all without requiring retraining or architectural changes.

Takeaway: A model-agnostic approach to the persistent problem of hallucinations in generative 3D/4D content, addressing both spatial duplication and temporal flicker in a unified way.

6. Multi-Perspective Agentic Program Repair via Code Property Graphs and Temporal Execution Graphs

Authors: Zhili Huang, Ling Xu, Hongyu Zhang | Categories: cs.SE, cs.AI Link: arxiv.org/abs/2607.12605v1

CT-Repair uses Code Property Graphs and Temporal Execution Graphs with three specialized agents analyzing bugs from static, dynamic, and hybrid perspectives. On 854 Java bugs from Defects4J, it correctly repairs 489 bugs in a mixed-model configuration and 388 with a controlled GPT-5.4-mini setup, outperforming ReinFix and RepairAgent. The filtering pipeline reduces runtime evidence by 94.85% on average.

Takeaway: Structured runtime evidence and multi-perspective reasoning significantly improve automated program repair without simply scaling up patch generation budgets.

7. ANGLE: Angular Neural Generative Learning via Engression

Authors: Rajdeep Pathak, Archi Roy, Tanujit Chakraborty | Categories: stat.ML, cs.LG, math.ST Link: arxiv.org/abs/2607.12833v1

ANGLE is a lightweight deep generative framework for circular/angular data regression that learns full conditional distributions via a generalized circular energy score. It handles multimodal and asymmetric distributions where traditional mean regression fails, with applications demonstrated in object pose estimation and wind direction prediction.

Takeaway: Fills an important gap in statistical machine learning for directional data, with theoretical guarantees and practical utility in autonomous vehicles and energy systems.

8. Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels

Authors: Roman Prosvirnin, Victor Minchenkov, Alexey Soldatov, Vladimir Bashun | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2607.12792v1

JADR measures a model’s internal danger recognition through Jacobian space analysis before generating any response, avoiding reliance on external judge models. Tested across six models and three quantization regimes (BF16, INT8, INT4), the protocol’s SafetyAUC metric statistically separates models with strong vs. weak internal safety mechanisms and captures different effects across quantization levels.

Takeaway: Probes beyond surface-level jailbreak testing to measure the hidden fragility of safety mechanisms in LLMs, revealing how quantization affects internal danger recognition in different ways.

This content was generated with AI assistance. Paper information sourced from arXiv.