Benchmarking LLMs Across Radar Sensing, Code Contamination, and Manipulation Control

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering advances in multimodal perception, AI security and auditing, dynamic robotic control, and foundational learning theory.

1. Can Language Models Understand mmWave Data? Benchmarking Large Language Models for mmWave Radar-Based Human Understanding

Authors: Jeongwan Shin, Jaehyeon Kim, Donguk Ko, Jaeho Choi | Categories: cs.AI Link: arxiv.org/abs/2608.14179v1

This paper introduces mmWave-QA, the first benchmark for language-conditioned human perception using millimeter-wave radar point clouds. The authors propose a minimal textualization interface that serializes radar data into natural language, enabling off-the-shelf LLMs to answer questions in a zero-shot setting. The benchmark spans six scenarios and five QA tasks, harmonizing heterogeneous public datasets with calibration-aware preprocessing.

Takeaway: A clever approach to bridging a non-visual sensing modality with LLM reasoning, potentially enabling robust perception in low-light or occluded environments where cameras fail.

2. Detecting Contaminated Code-Generation Prompt Batches via Influence Functions

Authors: Francesco Quinzan, Noor Munir, Yishun Lu, Stephen Roberts | Categories: cs.LG Link: arxiv.org/abs/2608.14303v1

The authors present CodeSIFT, a threat-model-agnostic method that uses influence functions to detect batches of prompts designed to elicit insecure code from LLMs. Instead of looking for known vulnerability patterns, CodeSIFT measures parameter-space influence and runs a statistical test to flag deviations from benign behavior. Evaluated on three open-weight code LLMs (3B–7B parameters), it achieves AUROC scores up to 0.98 and substantially outperforms static analysis baselines.

Takeaway: This is a promising direction for defending against novel and unknown attack classes, moving beyond signature-based defenses toward behavioral anomaly detection.

3. Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination

Authors: Shuo Liang, Yixing Ma, Pengfei Zhou, Xingyan Chen, Zihan Mei et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2608.14391v1

This paper introduces RA-Bench, a benchmark with 17,886 videos for evaluating AI-generated video detection, anchored by 1,830 real videos across 10 social-risk categories and complemented by 16,056 generated clips from nine generators. The systematic evaluation reveals that no detector family generalizes consistently, that videos which fool humans also fool detectors, and that social dissemination makes detection harder. The findings highlight a serious gap in current defenses against realistic synthetic video.

Takeaway: A sobering and comprehensive audit showing that current detectors are not robust enough for real-world crisis misinformation scenarios, underscoring the need for next-generation detection methods.

4. Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL

Authors: Xiaojun Wu, Cehao Yang, Honghao Liu, Xueyuan Lin, Zhichao Shi et al. | Categories: cs.CL Link: arxiv.org/abs/2608.14312v1

Envs-FORGE is a prompting policy that adaptively synthesizes RL training environments based on per-seed pass rates and a target learning frontier, using a mixed-integer linear program to select projection actions. The approach rewrites instructions, tests, and Docker environments with gold verification before training. On Qwen 3.5 35B, it improves Pass@1 by 9.2 points on tb-core and 6.4 points on tb-2.0 over the base model, exceeding fixed-recipe baselines.

Takeaway: A thoughtful departure from static few-shot or Evol-Instruct recipes, showing that dynamic, reward-grounded environment generation can meaningfully boost agent performance.

5. Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation

Authors: Yuxuan Chen, Wanruo Zhang, Xiao Li | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2608.14379v1

This paper introduces ReflexBench, a benchmark with six dynamic tasks for reaction-critical manipulation, and ReflexVLA, an efficient VLA model that improves temporal reasoning via latent future prediction and multi-frame temporal fusion. The model reduces deployment latency through batched visual encoding and CUDA Graph replay. Experiments show consistent improvements on dynamic tasks while maintaining competitive performance on static benchmarks, with real-world validation.

Takeaway: This work addresses a critical blind spot in robotics benchmarks—dynamic, reaction-critical scenarios—and offers a practical architecture that doesn’t require massive robot-data pretraining.

6. Whose doctor does the AI recommend? An algorithm audit of reputation and demographic signals in large language model-assisted physician choice

Authors: Syeda Anshrah Gillani, Mirza Samad Ahmed Baig | Categories: cs.CY, cs.AI, cs.CL Link: arxiv.org/abs/2608.14399v1

This prespecified randomized audit of seven LLMs (including gpt-4o-mini) examines what causally influences AI-assisted physician recommendations across 40,068 scored responses. The findings show that reputation signals dominate—a rating increase from 3.9 to 4.7 boosts choice probability by 31.4 pp—while demographic parity is rejected with female- and minority-signaled names gaining small but consistent advantages. Notably, models almost never mention these demographic drivers in their reasoning, rendering self-reported explanations unreliable for audit purposes.

Takeaway: A rigorous and repeatable audit methodology that demonstrates why external behavioral auditing—not model self-explanation—is necessary to catch subtle biases in consequential AI-mediated decisions.

7. A Hybrid LLM-Based Framework for Automated Security Annotation Generation in Business Process Models

Authors: Md Kamrul Islam, Tiphaine Henry, Mattia Salnitri, Julius Köpke, Sami Souihi | Categories: cs.CR, cs.AI, cs.SE Link: arxiv.org/abs/2608.14370v1

The authors present a hybrid framework that automatically generates SecBPMN2 security annotations from BPMN process models and natural-language security requirement documents, combining LLM-based extraction with schema-constrained mapping and rule-based validation. Evaluated on 27 process models, it achieves substantially higher precision than human analysts (0.58 vs. 0.29) with comparable recall, reduces erroneous annotations by nearly 50%, and is significantly faster.

Takeaway: A strong example of how LLM-plus-rule hybrid systems can outperform pure human effort or pure LLM approaches in structured, high-stakes domains like security-by-design modeling.

8. Online Inference in Distributional Temporal-Difference Learning

Authors: Yang Peng, Liangyu Zhang | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2608.14408v1

This paper establishes theoretical guarantees for online statistical inference on return distributions in reinforcement learning, using nonparametric distributional temporal-difference learning. The authors prove asymptotic normality of the averaged estimator in Cramér space and show that a bootstrap analogue converges to the same limit, enabling inference for smooth functionals like variance and CVaR. A local asymptotic theory is also developed for nonsmooth functionals such as quantiles.

Takeaway: This foundational theory paper provides the missing statistical machinery for uncertainty quantification in distributional RL, bridging a key gap between practice and rigorous inference.

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