Adaptive RL Intervention, Long-Video Memory Routing, and Agent Security

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering adaptive intervention for real-world robot RL, LLM-driven attack path generation, compositional memory routing for interactive video, confidence-guided multimodal RAG, dual-loop mobile manipulation, a unified theory of validation and information fusion, a critical re-evaluation of prompt-injection detectors, and input representations for multi-pitch estimation in vocal ensembles.

1. UniIntervene++: An Adaptive Intervention Agent for Efficient Real-World Reinforcement Learning

Authors: Yudong Lin, Haoyuan Deng, Zhuoxuan Yuan, Zaijia Yang, Yuanjiang Xue et al. | Categories: cs.LG, cs.RO Link: arxiv.org/abs/2610.03620

UniIntervene++ is an adaptive intervention agent that learns to allocate control between autonomous execution and heterogeneous assisted behaviors during online reinforcement learning. It unifies the evolving RL policy, trajectory correction, and a task-structured CodePolicy as Options in a semi-Markov decision process, then uses competence-adaptive probing and coupled experience learning to jointly decide when to intervene, how to intervene, and when to return control. Across five real-world manipulation tasks, it reaches an average success rate of 89.67% (at least 6 points above all baselines) while cutting human intervention to 0.77%, a relative reduction of at least 94.6%.

Takeaway: A rare example of intervention scheduling that adapts as the policy itself improves, with the human-effort savings being the headline result for anyone deploying RL on physical robots.

2. CVE2AP: Automated Generation of PDDL-Encoded Attack Paths via Large Language Models

Authors: Lin Cui, Vincenzo Scotti, Raffaela Mirandola | Categories: cs.AI Link: arxiv.org/abs/2610.03383

Attack path modeling with PDDL is typically built by hand and struggles to keep pace with evolving threats, so CVE2AP uses LLMs with structured prompting to generate PDDL-encoded attack paths directly from natural-language CVE descriptions. An error-feedback loop iteratively refines outputs using planner-reported syntactic and solvability errors. Empirical evaluation across multiple LLMs and configurations reaches up to 86.9% syntax correctness, 78.6% solvability, and 93.1% semantic correctness, with GPT-5.5 offering the best quality-cost trade-off and error feedback giving the most consistent gains.

Takeaway: A practical demonstration that planner-in-the-loop feedback, not just bigger models, is what makes formal security artifacts generatable from unstructured threat text.

3. Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation

Authors: Ziyi Wang, Junchi Yao, Heqian Qiu, Wenbo Shi, Chengjiu Wang et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2610.03510

Interactive long-video generation needs new shots that recombine characters and backgrounds from different historical moments, which whole-prompt retrieval cannot handle and naive memory mixing corrupts with unrelated content. Weave Forcing is a training-free framework that uses LLM semantic slot routing to decompose prompts into character and background components, then applies masked memory weaving with contrasting attention maps to expose only relevant tokens from compressed historical KV memories. Coverage adaptive RoPE further adjusts temporal offsets and memory retention depending on no, partial, or full reference coverage, improving cross-shot subject and background consistency at competitive visual quality.

Takeaway: Component-level memory routing is a clean answer to the “which past shot do I borrow from” problem, and doing it training-free makes it easy to bolt onto existing autoregressive video stacks.

4. CLIMB: Confidence-Guided Complementary Evidence for Multimodal Retrieval-Augmented Generation

Authors: Hang Gao, Wujiang Xu, Zhixing Zhang, Kai Mei, Jingyi Yang et al. | Categories: cs.CL Link: arxiv.org/abs/2610.03421

Multimodal RAG systems relying on Top-K retrieval or reranking often surface redundant passages and give little control over whether an answer update is actually supported by the evidence. CLIMB is a training-free inference-time framework that builds a compact complementary evidence pool via an MMR-style objective balancing relevance and redundancy, then refines answers within that fixed pool using an R/E/C critic and an evidence-grounded confidence estimator that accepts updates only when confidence rises. Experiments on Encyclopedic-VQA and InfoSeek show consistent gains over retrieval-augmented multimodal baselines, with ablations confirming each component contributes.

Takeaway: The confidence-based stopping criterion is the quietly valuable part — it gives multimodal RAG a principled way to decline further refinement instead of iterating blindly.

5. MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation

Authors: Chenzhi Liu, Yue Zhang, Jiehong Lin, Jianan Wang, Bo Wang et al. | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2610.03476

Long-horizon mobile manipulation suffers from compounding execution errors and interference between locomotion and arm control, and existing hierarchical agents use rigid sub-task mappings with no continuous learning. MobiAgent pairs an Inner Loop, where a VLM performs receding-horizon planning and visual reflection over composable atomic skills executed by flow-matching experts sharing a unified VLM backbone, with an Outer Loop that autonomously segments, verifies, and clusters deployment rollouts to discover and fine-tune skills without human annotation. It outperforms π₀.₅-TA by 22.5 points on BEHAVIOR-1K, and autonomous data recycling lifts success from 7.50% to 27.50% on RoboCasa and 32.5% to 57.5% on Astribot S1.

Takeaway: The outer loop’s autonomous rollout mining is the key idea — it turns deployment failures into training data, which is exactly what long-horizon manipulation needs to escape manual data collection.

6. When Is Accuracy Evidence? A Unified Theory of Generalisation, Validation, and Information Fusion

Authors: JM Gorriz | Categories: stat.ML, cs.LG, physics.data-an Link: arxiv.org/abs/2610.03465

This work generalizes Cross Upper-Bound Validation (CUBV) into a single exponential framework where the moment-generating function of the generalization gap is controlled by a cumulant envelope γ(λ), yielding risk bounds that span Hoeffding, Bernstein, dependency-aware, PAC-Bayesian, and heterogeneous source-fusion settings. Modeling fold dependence through a joint sub-Gaussian proxy matrix gives an effective number of folds K_eff = K/[1+(K−1)ρ], showing that raising K does not necessarily add statistical evidence when folds are strongly correlated. Experiments on heterogeneous multimodal Gaussian mixtures show that in low-dimensional small-sample regimes K-fold partitioning can inflate uncertainty, while corrected resubstitution remains valid and tighter — an effect that vanishes as sample size grows.

Takeaway: Worth reading for the K_eff result alone: it formalizes the intuition that more folds can mean less evidence, and gives small-sample practitioners a concrete validation alternative.

7. Passing the Test You Trained On: Re-evaluating Prompt-Injection Detectors for LLM Agents

Authors: Zhuowen Liu | Categories: cs.CR, cs.CL Link: arxiv.org/abs/2610.03448

This paper asks whether benchmark scores predict how a prompt-injection detector actually behaves inside an LLM agent, replaying ground-truth tool calls from AgentDojo and tau-bench without an LLM to produce benign-by-construction tool outputs, labeling injections by differential replay, and evaluating fifteen detectors plus two LLM judges. Detection rankings transfer poorly across benchmarks — the best BIPIA detector catches 2% of AgentDojo injections at a 1% false-positive rate, and a detector catching 72% on AgentDojo catches 15% on tau-bench — though false-positive rates do transfer. Training input form explains the gap: the detector that wins on both agent benchmarks shares no data with either and was trained on agent-style inputs.

Takeaway: A sobering and necessary result — if you pick a prompt-injection detector off a leaderboard, you are probably measuring the wrong thing; evaluate on your own tool outputs at a low false-positive rate.

8. Revisiting Input Time-frequency Representations in Multi-pitch Estimation for Vocal Ensembles

Authors: Junyoung Koh, Hao-Wen Dong | Categories: cs.SD, cs.AI Link: arxiv.org/abs/2610.03656

Multi-pitch estimation in vocal ensembles is hard because overlapping pitch ranges cause harmonics to collide in time-frequency representations, and the common harmonic constant-Q transform (HCQT) is expensive to extract when training mixtures are generated on the fly. This paper compares HCQT against a linear short-time Fourier transform whose frequency bins feed the model directly, finding that the linear STFT outperforms HCQT while substantially reducing feature-extraction cost. Further analysis shows longer analysis windows or broader spectral coverage add nothing, and restricting input to the predicted pitch range erodes the linear STFT’s advantage.

Takeaway: A useful reminder that frequency-adaptive front ends are not automatically better, and that shorter analysis windows suit time-varying vocal pitches.

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