Agentic Black-Box Optimization, Recursive Self-Improvement, and Multimodal World Modeling

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering agentic black-box optimization benchmarks, recursive self-improvement frameworks, visual world modeling for multimodal LLMs, alignment generalization prediction, robot foundation model reasoning, causal data fusion, recursive safety improvement, and audio encoder inversion audits.

1. A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization

Authors: Ming Chen, Rong-Xi Tan, Ke Xue, Yu-Jie Zhou, Taiye Lu et al. | Categories: cs.LG, cs.AI, cs.NE Link: arxiv.org/abs/2610.12183

Existing studies of LLM agents for black-box optimization (BBO) use inconsistent task domains and configurations, making results hard to compare. The authors introduce AgenticBBO-Bench, a cross-domain benchmark spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget protocol. Agentic BBO outperforms direct LLM methods in all five domains and beats the best numerical optimizers in four, with a five-task frontier challenge showing GPT-6 Astra and DeepSeek-V4.1-Flash on the performance-cost Pareto frontier.

Takeaway: A much-needed standardization effort for the fast-growing but fragmented field of LLM-driven optimization, with honest findings about which design choices actually matter.

2. Recursive Self-Improvement through Multi-Agent Self-Supervision

Authors: Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi et al. | Categories: cs.AI Link: arxiv.org/abs/2610.12176

Recursive self-improvement on non-verifiable tasks hits a supervision bottleneck where the model itself becomes the best available optimizer and evaluator. The authors propose Multi-Agent Self-Supervision (MASS), which alternates between evolutionary workflow optimization and supervised fine-tuning on self-generated multi-agent trajectories. Over two cycles with Qwen3.6-27B, MASS achieves 1.2–1.6x higher performance per output token on four open-ended benchmarks, with multi-agent traces proving more training-efficient than single-agent data.

Takeaway: A concrete recipe for bootstrapping model capabilities without human supervision, and the finding that orchestration itself can be learned is a notable step toward practical RSI.

3. WOVEN: Weaving Visual World Modeling into Multimodal LLMs

Authors: Zheyu Fan, Yue Zhang, Mingkai Deng, Kangrui Wang, Qineng Wang et al. | Categories: cs.CV, cs.CL, cs.LG Link: arxiv.org/abs/2610.12417

MLLMs struggle with spatial, embodied, physical, and temporal reasoning — deficits the authors hypothesize stem from a shared weakness in visual transition reasoning. They introduce WOVEN, a training source and benchmark with 36,076 examples across 20 scene types, 5 action types, and 8 reasoning types, generated from video-pretrained models. After evaluating 38 frontier MLLMs and finding systematic deficits, they show that training on WOVEN transfers broadly, improving 22 of 26 external benchmarks by up to 27.3 points.

Takeaway: Reframes a cluster of MLLM failures as one learnable primitive, and the resulting training recipe — select supervision by reasoning operation, not domain — is broadly actionable.

4. Predicting Alignment Generalization with Value Representations

Authors: Andy Liu, Mehar Bhatia, Karolina Stanczak, Mona Diab, Vered Shwartz et al. | Categories: cs.CL, cs.AI, cs.LG Link: arxiv.org/abs/2610.12410

Post-training on narrow behaviors can unexpectedly shift model behavior in unseen contexts, yet predicting these spillovers remains an open problem. The authors formalize alignment generalization prediction, analyzing effects across 66 values and benchmarking representational techniques. Activation-based representations achieve correlations of 0.45 with their generalization matrix versus 0.05 for description-based baselines, and they use these representations to sketch the first taxonomy of LLM values grounded in empirical generalization dynamics.

Takeaway: Turns an understudied alignment failure mode into a measurable prediction task, and the striking gap between activation- and description-based methods has clear implications for alignment target design.

5. ARC: A Reasoning Recipe for Robot Foundation Models

Authors: Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss, Zhaoshuo Li et al. | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2610.12386

Rather than scaling robot foundation models with more data and compute, the authors show that a reasoning recipe — grounded traces explaining why an action is appropriate and what effect it should produce — can substantially improve zero-shot performance. These traces are generated automatically from existing demonstrations, yielding ARC-Trace-DROID without new robot data, and can be learned by VLAs such as π₀.₅ and WAMs such as Cosmos3-Nano-Policy. The adapted models set new state of the art on RoboLab-120 and MolmoSpaces, with up to 50-point gains on RoboLab-Reasoning-50 and 82.2-point real-robot improvement for π₀.₅.

Takeaway: A compelling counterpoint to brute-force scaling in robotics, with causal action traces as a lightweight, data-efficient lever that delivers outsized gains.

6. Prediction-Powered Data Fusion for Treatment Effect Estimation

Authors: Yonghan Jung, Shu Yang | Categories: stat.ML, cs.AI, cs.LG, stat.ME Link: arxiv.org/abs/2610.12332

Combining small RCTs with large observational studies is attractive but existing estimators either impose strong assumptions on the observational data, fail to borrow enough power, or trade bias for variance — and CATE has received less attention than ATE. The authors propose a fusion framework that preserves RCT-based unbiasedness while borrowing precision from large observational data without special assumptions. Applying this principle yields AIPW-Fusion for ATE with closed-form weights and confidence intervals, plus DR-Fusion and R-Fusion for CATE.

Takeaway: A principled unification of prediction-powered inference with causal effect estimation, offering practical estimators with both theoretical guarantees and finite-sample results.

7. ReSI: Recursive Safety Improvement toward Resistant and Resilient AI

Authors: Jingnan Zheng, Dongcheng Zhang, Yi Zhang, Ming Zhang, Qiaosheng Zhang et al. | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2610.12233

As models evolve through frequent updates, safety alignment must adapt to each new checkpoint while generalizing to risks not yet revealed — goals the authors term resistance and resilience. ReSI applies recursive self-improvement to safety itself: each round runs diverse red-teaming, develops training recipes, and promotes the update with the largest safety gain among those passing a capability-retention Pareto gate. Across four dense and MoE models, ReSI cuts mean X-Teaming attack success from 86.01% to 31.45%, below GPT-5.6-Luna’s frontier result of 56.69%, while largely preserving capabilities.

Takeaway: Flipping recursive self-improvement into a defensive tool is a smart inversion, and the out-of-distribution resilience numbers suggest the approach generalizes beyond exposed vulnerabilities.

8. How Much Audio Is Left In An Embedding? An Inversion Audit Of Audio Encoders

Authors: Marios Glytsos, Brian McFee | Categories: cs.SD, cs.AI Link: arxiv.org/abs/2610.12250

Pretrained audio encoders are reused for downstream tasks unknown at training time, so their value depends on which signal properties survive the pretext objective. Using a shared Stable Audio Open latent diffusion decoder, the authors reconstruct five-second stereo music from frozen representations of supervised classifiers, a contrastive model, and a reconstruction model. Evaluations on the Million Song Dataset reveal clear reconstructability differences across encoder families, with finer temporal or spectral interfaces aiding recovery — and even compressed task-oriented embeddings retain measurable source specificity.

Takeaway: An informative audit with privacy implications: task-oriented audio embeddings leak more recoverable content than their compression suggests, and interface resolution matters as much as objective.

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