LLM Honesty Training, Agent Security Guards, and Personalized Robot Manipulation

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering honesty alignment for LLMs, agentic control of radio networks, graded multimodal verifiers, 3D code generation, personalized robot manipulation benchmarks, closed-loop neural control, multilingual security of agent guards, and LLM-discovered quantum circuits.

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

The authors introduce CATCH (Continuous Adversarial Training for Consistent Honesty), which trains models toward honest answers and away from dishonest ones using only hundreds of self-reporting targets, while preserving capabilities. Across Llama-3.3-70B-Instruct and Qwen3-3.6-27B, CATCH reduces dishonesty under pressure on MASK, increases confessions in Among Us, and improves hidden-behavior discovery in AuditBench. It outperforms prompting and supervised fine-tuning, matches honesty activation steering, and surpasses steering on utility thanks to a novel utility-data choice.

Takeaway: A rare honesty intervention that generalizes and transfers across benchmarks without sacrificing capability—worth reading for anyone working on alignment beyond surface-level prompting.

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

X-RIC integrates agentic AI into O-RAN architectures, combining reactive reinforcement learning with a Telecom Language Model for contextual abstraction and interpretability, plus closed-loop co-adaptation in the Near-RT RIC. The authors also propose BARX, a domain-specific benchmark assessing latency, compliance, and determinism. In SLA-aware network slicing, X-RIC achieves 89% operational cost reduction with SLA compliance, converging 75% faster using 50% fewer training steps and 46% lower variance.

Takeaway: A concrete blueprint for pairing LLM reasoning with RL in telecom control loops—the BARX evaluation framework alone makes it valuable for practitioners deploying agentic systems in latency-sensitive domains.

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, this work introduces a verifier that decomposes complex prompts or environmental rules into fine-grained components and assigns pointwise scores for continuous, interpretable critique signals. A decoupled RL training strategy jointly optimizes sub-question-level binary judgment and global alignment scoring. Experiments show pointwise graded verifiers significantly outperform binary ones in text-to-image tasks, while theoretical and empirical analysis on Maze and Sudoku shows pairwise verification is more accurate when ground-truth images exist.

Takeaway: The pointwise-versus-pairwise distinction is a genuinely useful design insight for anyone building reward models or verifiers for structured world-modeling agents.

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 a research ecosystem for coding agents to create editable, verifiable 3D content, comprising 3DCodeVerse-2M (a large curated dataset of executable 3D code) and a multi-agent DAG harness that uses execution checks, spatial measurements, and visual feedback to automatically evaluate and revise programs. Fine-tuning on the dataset significantly improves open-source models’ 3D coding, and the harness achieves a 90% win rate against Claude Design in pairwise human evaluations when both are driven by Claude Opus 5.5.

Takeaway: Execution-grounded verification is emerging as the key to reliable 3D generation—this paper offers both the data and the harness architecture to pursue it.

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

LIBERO-PeRM is a controlled benchmark for preference learning in embodied agents that separates task completion from preference satisfaction, studying zero-shot preference following, learning from demonstrations, sample efficiency, generalization across layouts, and preference composition under conflicts. Built on LIBERO, it provides five task suites (312 tasks) with paired demonstrations and preference-satisfaction metrics. Experiments reveal current VLA policies struggle with zero-shot personalization; SFT makes preferences learnable, and mixed training improves retention, but generalization, decomposition, and conflict handling remain difficult.

Takeaway: Personalization is the next frontier for household robots, and this benchmark finally gives the field a rigorous way to measure it—expect it to become a standard testbed.

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

This paper argues that offline-trained neural networks deployed in closed-loop control fundamentally alter the state distribution they were trained on, and introduces a taxonomy of sequential failure modes: Closed-Loop Identification Error, deployment-induced Action Error, and continuous Tracking Error. Discrete-time sensitivity analysis formalizes how offline-learned parameters guarantee trajectory divergence upon deployment, and ablation studies on a nonlinear Duffing oscillator show that time-scale-separated online adaptation mitigates divergence where static deployment fails.

Takeaway: A crisp theoretical framing of a problem practitioners often hit empirically—essential reading for anyone deploying learned components in safety-critical feedback systems.

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

The authors build LinguaGuardBench from 1,220 labelled steps of TS-Bench AgentDojo-Traj, translating only the user instruction into German, French, Turkish, Swahili, and Amharic while keeping all tool calls and labels identical. Across 19 guards, most small, open, and purpose-built guards change verdicts under translation beyond their noise floor, with the mean placebo-adjusted flip rate rising from +1.5 pp on French to +4.4 pp on Amharic; purpose-built guards are most sensitive, while frontier guards shift by at most about one point. The benchmark and all 489k verdicts are released.

Takeaway: A sobering demonstration that language itself is an under-defended attack surface for agent safety—multilingual robustness should now be a baseline requirement for guard models.

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 present 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 relative to state of the art; in the eighth, the search finds a new reversible architecture for prime-field modular inversion that makes divsteps viable by handling reversibility differently.

Takeaway: Evaluator-guided LLM search producing genuinely novel algorithmic structures—not just parameter tweaks—is a compelling signal that AI-driven scientific discovery is maturing.

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