LLM Honesty, Robotic Personalization, and Quantum Circuit Discovery

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering honesty alignment in LLMs, agentic control of Open RAN networks, multimodal verifiers, 3D code generation, personalized robotic manipulation, closed-loop neural control, multilingual security vulnerabilities in agent guards, and LLM-driven quantum circuit discovery.

1. CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty

Authors: Yuan-Chen Chang, Niklas Herbster, Gauthier Gidel, Tommaso Tosato | Categories: cs.LG, cs.CL, cs.CY Link: arxiv.org/abs/2610.22292

Despite alignment training, LLMs still exhibit dishonest behavior such as contradicting known facts or making claims unsupported by context. The authors introduce CATCH (Continuous Adversarial Training for Consistent Honesty), which trains models toward honest answers and away from dishonest ones while preserving capabilities. Tested on Llama-3.3-70B-Instruct and Qwen3-3.6-27B with only hundreds of self-reporting targets, CATCH reduces dishonesty under pressure, increases confessions in Among Us, and improves hidden-behavior discovery in AuditBench — outperforming prompting and SFT, and matching honesty activation steering while better preserving utility.

Takeaway: A rare alignment technique that improves honesty without sacrificing capability, with demonstrated transfer across benchmarks and architectures. Worth reading for the utility-data design alone.

2. X-RIC: An Agentic RAN-Intelligent Controller Benchmark for Enhancing Reinforcement Learning

Authors: Adhwaa Al-Chaab, Ayman Younis, Dario Pompili | Categories: cs.AI, cs.AR Link: arxiv.org/abs/2610.22269

Current AI approaches to O-RAN’s RAN Intelligent Controller lack reasoning and explainability. X-RIC integrates agentic AI via automated behavioral cloning, a Telecom Language Model for contextual reasoning, and closed-loop co-adaptation in the Near-RT RIC. The authors also introduce BARX, a benchmark for agentic RAN reasoning, and validate X-RIC on SLA-aware network slicing — achieving 89% operational cost reduction, 75% faster convergence, 50% fewer training steps, and 46% lower variance.

Takeaway: A concrete demonstration that agentic reasoning can be folded into telecom control loops with measurable efficiency gains. The BARX benchmark is a useful contribution for the O-RAN research community.

3. OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling

Authors: Bowei Liu, Xinchen Zhang, Xuhuan Li, Tianlin Zhang, Kaian Jiang et al. | Categories: cs.CL Link: arxiv.org/abs/2610.22355

Moving beyond binary True/False verifiers, the authors introduce a verifier that decomposes complex prompts or environmental rules into fine-grained components and assigns pointwise scores, yielding continuous and interpretable critique signals. A Decoupled RL training strategy jointly optimizes sub-question-level binary judgment and global alignment scoring. The pointwise graded verifier substantially outperforms binary verifiers in text-to-image tasks and extends effectively to rule-governed world modeling (e.g., Maze, Sudoku), with theory showing pairwise verification is more accurate when ground-truth images exist.

Takeaway: The shift from binary to graded verification is a meaningful step toward reliable multimodal critics. The insight about pairwise vs. pointwise verification depending on ground-truth availability is particularly valuable.

4. 3DCodeVerse: Exploring 3D Worlds as Code

Authors: Yipeng Gao, Ziyao Zeng, Jinfa Huang, Yijiang Li, Haolin Xiong et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2610.22341

3DCodeVerse is an ecosystem for coding agents to create editable, verifiable 3D content at scale, comprising a 2M-example dataset of executable 3D code, a multi-agent DAG harness for generating and refining 3D programs via execution checks and visual feedback, and systematic empirical studies. Fine-tuning on the dataset significantly improves open-source models’ 3D coding, and the harness achieves a 90% win rate against Claude Design in human pairwise evaluations when both are driven by Claude Opus 5.5.

Takeaway: Treating 3D content as executable code — with automated verification — is a compelling alternative to direct generative approaches. The 90% win rate against a strong baseline is a striking result.

5. LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

Authors: Zhixu Li, Keqian Tang, Litian Gong, Jingyu Yao, Wenqian Zhang et al. | Categories: cs.RO, cs.AI, cs.LG Link: arxiv.org/abs/2610.22372

Personalization is essential for household robots, yet personalized robotic manipulation lacks a controlled benchmark separating task completion from preference satisfaction. LIBERO-PeRM studies five questions about preference learning — zero-shot following, learning from demonstrations, data efficiency, generalization, and composition under conflict — with 312 tasks, paired demonstrations, and preference metrics. Experiments show current VLA policies struggle with zero-shot personalization, SFT makes preferences learnable, but generalization, decomposition, and conflict handling remain difficult.

Takeaway: A well-structured benchmark that exposes a real gap between task completion and preference satisfaction in embodied agents. Essential for anyone working on personalized robotics or preference learning.

6. Learning Changes the World and Then What Was Learned has Changed: Training and Deploying Neural Networks in the Control Loop

Authors: Vittorio Lippi | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2610.22084

When offline-trained neural networks are deployed in closed-loop control, their actions alter the state distribution they were trained on, causing sequential failure modes. The paper introduces a taxonomy of these failures — Closed-Loop Identification Error, deployment-induced Action Error, and continuous Tracking Error — formalized via discrete-time sensitivity analysis. The author proposes time-scale-separated online adaptation and validates it on a nonlinear Duffing oscillator, showing bounded continuous learning mitigates trajectory divergence where static deployment fails.

Takeaway: A rigorous formalization of a widely felt but under-theorized problem in ML-for-control. The concrete design guidelines for safety-critical feedback systems make this practically useful.

7. Language Is a Security Surface for Agent Guards

Authors: Hasan Demirkiran, Jens Ernstberger | Categories: cs.CR, cs.AI, cs.CL, cs.LG Link: arxiv.org/abs/2610.22244

Step-level agent guards are trained and evaluated in English, but users write requests in many languages. Building LinguaGuardBench from 1,220 labeled steps with instructions translated into German, French, Turkish, Swahili, and Amharic while keeping tool calls and labels byte-identical, the authors find that most open, small, and purpose-built guards change verdicts under translation beyond their noise floor — with flip rates rising from +1.5 pp on French to +4.4 pp on Amharic among sub-13B guards. Frontier guards shift by at most about one point, and the damage concentrates on steps where reasoning restates the English request.

Takeaway: A sharply designed study exposing language as a genuine attack surface for agent safety infrastructure. The release of 489k verdicts makes this immediately actionable for guard developers.

8. Discovering state-of-the-art quantum circuits with large language models

Authors: Andi Gu, Alek H. Kemeny, Susanne F. Yelin, Mikhail D. Lukin | Categories: quant-ph, cs.AI Link: arxiv.org/abs/2610.22367

The authors introduce an evolutionary framework in which LLMs propose quantum circuit designs and an independent evaluator verifies and scores them, applied across numerical computation, chemistry, optimization, quantum simulation, and number theory. In seven of eight tasks, discovered circuits significantly reduce both Toffoli and qubit counts versus the state of the art; the eighth discovers a new reversible architecture for prime-field modular inversion that makes divsteps — previously impractical in reversible arithmetic — viable.

Takeaway: A powerful demonstration that evaluator-guided LLM search can produce genuinely novel algorithmic structures, not just parameter tuning. The reversibility-aware architecture discovery is a standout result.

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