LLM Agents for Optimization, Recursive Self-Improvement, and Multimodal World Modeling

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

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

Black-box optimization is ubiquitous in science and engineering, and LLM agents offer a promising new approach by combining task semantics, computational tools, and feedback-driven decision making. To address the incomparability of existing agentic BBO studies, 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. Experiments show agentic BBO outperforms direct LLM methods in all five domains and beats the best numerical optimizers in four, while ablations reveal that extra numerical tools don’t consistently help, task semantics are broadly useful, and specific priors are less reliable.

Takeaway: The first rigorous, apples-to-apples benchmark for agentic black-box optimization—its negative results (more tools ≠ better) are as valuable as the leaderboard itself.

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 when model outputs exceed human expert assessment, leaving the model itself as 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 trajectories, prompting a single base model to iteratively propose, execute, and self-evaluate multi-agent workflows. Over two cycles with Qwen3.6-27B, MASS achieves 1.2–1.6× higher performance per output token on four open-ended benchmarks, and multi-agent traces prove more training-efficient than single-agent alternatives.

Takeaway: A concrete demonstration that orchestration—not just raw capability—can be bootstrapped recursively, with multi-agent traces emerging as a denser training signal than single-agent ones.

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

Multimodal LLMs struggle with spatial, embodied, physical, and temporal reasoning, which the authors hypothesize stems from a shared deficit in visual transition reasoning. They introduce WOVEN, a training source and benchmark of 36,076 examples across 20 scene types, 5 action types, and 8 reasoning types built from video-pretrained generative model rollouts. Evaluating 38 frontier MLLMs reveals systematic deficits even at scale, yet training on WOVEN subsets of only ~2,000 items collectively improves 22 of 26 external benchmarks by up to 27.3 percentage points—with the recipe favoring supervision selected by reasoning operation and larger visual state changes.

Takeaway: Establishes visual transition reasoning as a reusable training primitive, plus a prospectively validated recipe for how to select world-modeling supervision.

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-trained models score highly on alignment evaluations, yet training on narrow behaviors still influences behavior in unexpected ways across unseen contexts. The authors formalize alignment generalization prediction—forecasting how fine-tuning on one value changes behavior across held-out values—and conduct a large-scale analysis across 66 values. Activation-based representations significantly outperform textual descriptions (correlation of 0.45 vs. 0.05), and the authors use these representations to measure value similarity within multi-value targets, correlate it with robustness, and sketch a first empirically grounded taxonomy of LLM values.

Takeaway: A rare empirical handle on a notoriously slippery problem—value generalization—with a clear methodological verdict: measure values through activations, not descriptions.

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 larger models and more demonstrations, the authors show that a reasoning recipe—ARC—can substantially improve zero-shot task performance. ARC consists of reasoning traces grounded in the robot’s next action and its causal structure, a scalable automatic labeling pipeline that generates traces from existing demonstrations (producing ARC-Trace-DROID from DROID without new robot data), and adaptation strategies for VLAs like π₀.₅ and WAMs like Cosmos3-Nano-Policy. The adapted models set new state of the art on RoboLab-120 and MolmoSpaces with gains up to 50 points on RoboLab-Reasoning-50, and improve π₀.₅’s real-robot task success by 82.2 percentage points.

Takeaway: A strong data-and-reasoning-first counterpoint to brute-force scaling in robotics, with unusually large real-robot gains from zero new demonstrations.

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 randomized controlled trials with large observational studies is attractive but existing estimators either impose assumptions on the observational data or fail to borrow enough power from it. The authors propose a fusion framework that preserves the unbiasedness of RCT-based estimation while borrowing precision from the large OBS without special assumptions on it. Applying this principle yields AIPW-Fusion for the ATE with closed-form weights and confidence intervals, plus two CATE learners, DR-Fusion and R-Fusion.

Takeaway: A clean “no free lunch required” formulation for RCT-OBS fusion—closed-form ATE weights and confidence intervals make it immediately practical for applied causal work.

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 continually adapt—mitigating exposed vulnerabilities (resistance) while generalizing to risks not yet revealed (resilience). ReSI implements recursive safety improvement through automated research: each round applies diverse red-teaming methods, develops training recipes, and promotes the update with the largest safety gain among those passing a Pareto gate on capability retention. Across four dense and MoE models, ReSI matches or exceeds frontier models on in-distribution and out-of-distribution safety benchmarks, reducing mean X-Teaming attack success from 86.01% to 31.45%—well below GPT-5.6-Luna’s 56.69%—while largely preserving general capabilities.

Takeaway: Turns recursive self-improvement into a safety asset rather than just a risk, with a Pareto gate that gives a principled answer to the safety-capability tradeoff.

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 usefulness depends on which signal properties survive the pretext objective. The authors perform a paired source reconstruction audit, using a shared Stable Audio Open latent diffusion decoder to reconstruct five-second, 44.1-kHz stereo music from frozen representations of VGGish, ConvNeXt, CLAP, and EnCodec. Evaluating on the Million Song Dataset reveals clear differences in reconstructability across encoder families, with improved recovery when finer temporal or spectral structure is exposed—even compressed task-oriented embeddings retain measurable source specificity and high-level musical content.

Takeaway: An important privacy- and utility-relevant audit showing that task-oriented audio embeddings leak far more of their source than one might assume—and that interface granularity shapes how much survives.

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