Efficient LLM Inference, OLED Design, and Quantum Kernel Robustness

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering advances in domain-specific LLMs, cost-efficient inference, robotic learning, quantum computing, and low-resource OCR.

1. OLEDLM: A Unified Language Model for OLED Molecular Design

Authors: Fukang Wen, Yuchong Tang, Jingyuan Li, Beichen Wang, Yixuan Jiang et al. | Categories: cs.LG Link: arxiv.org/abs/2607.20194v1

This paper proposes an inverse molecular design framework based on causal language models that generates OLED SMILES sequences directly from target optoelectronic properties. Using a multi-stage strategy including a foundational LLaMA-style transformer, BERT-based property predictors, and Reinforcement Learning with DFT verification, the model efficiently navigates the vast OLED chemical space. This represents the first successful adaptation of LLMs specifically for the OLED domain.

Takeaway: A compelling validation that LLMs can be adapted for highly specialized scientific discovery tasks where labeled data is scarce and quantum-chemical constraints are stringent.

2. PyroDash: Cost-Efficient Token-Level Small-Large Language Model Collaborative Inference

Authors: Niqi Lyu, Pengtao Shi, Wei Qiu, Jianlin Zhong, Sicong Xia et al. | Categories: cs.CL Link: arxiv.org/abs/2607.20327v1

This paper introduces PyroDash, a framework where a small language model can request assistance from a large language model at the token level via learned control tokens, requiring no separate router or LLM retraining. Training uses a three-stage process with cost-aware alignment via Group Relative Policy Optimization. On mathematical reasoning benchmarks, it achieves 64% accuracy (6.36 points above LLM-only) while reducing cost by 20.4%, or at an aggressive cost setting, cuts total inference cost from $49.36 to$1.78.

Takeaway: A practical and impressive demonstration of how token-level collaboration between small and large models can dramatically reduce inference costs while preserving—or even improving—reasoning accuracy.

3. PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning

Authors: Alexis Fox, Junlin Wang, Paul Rosu, Bhuwan Dhingra | Categories: cs.AI Link: arxiv.org/abs/2607.20064v1

This paper proposes PRO-LONG, a context management framework using programmatic memory that keeps a complete structured interaction log and uses coding agents to search history efficiently for long-horizon LLM reasoning tasks. On ARC-AGI-3, it improves over base coding agents by 18 percentage points across frontier models, matching state-of-the-art specialized harnesses (up to 76.1% pass@1) while using 4.2-5.8x fewer tokens. With Fable 5, PRO-LONG achieves 97.4% best@2 at a total cost of $1,750.

Takeaway: A clever solution to the memory-management tradeoff in long-horizon agent tasks, demonstrating that structured programmatic memory can dramatically reduce token usage while improving performance.

4. Persian Pixel: A large-scale synthetic OCR dataset for Persian language

Authors: Pouria Mahdi, Haq Nawaz Malik | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2607.20385v1

This paper introduces Persian Pixel, a synthetic OCR dataset with over 343,000 high-fidelity Persian image-text pairs generated from a seven-million-word corpus. The pipeline models complex Persian script features including cursive connectivity, context-dependent glyph shaping, and ligatures, with more than 25 stochastic degradation models to bridge the synthetic-to-real gap. The dataset provides a scalable resource for training transformer-based OCR models like TrOCR and Donut.

Takeaway: A timely contribution that addresses a critical data bottleneck for Persian OCR, offering a cost-effective synthetic alternative to manual annotation for a typographically complex, low-resource script.

5. Statistical Inference for Rank Allocation in Low-Rank Adaptation

Authors: Yihang Gao, Vincent Y. F. Tan | Categories: stat.ML, cs.LG, math.ST Link: arxiv.org/abs/2607.20205v1

This paper proposes StatLoRA, which formulates LoRA rank allocation as a statistical hypothesis testing problem, using p-values to determine which components to retain or prune. The method is supported by a central limit theory for stochastic optimizer trajectories, establishing asymptotic normality for optimizers including AdamW. Experiments on DeBERTaV3-base, BART-Large, and Qwen2.5-7B show comparable or better performance than vanilla LoRA and adaptive methods under matched rank budgets.

Takeaway: A principled statistical foundation for an important practical problem in parameter-efficient fine-tuning, providing theoretical guarantees for adaptive rank allocation.

6. Small, Free, and Effective: Orchestrating Open-Weight Small Language Models to Outperform Single LLM for Malware Analysis

Authors: Adel ElZemity, Shujun Li, Budi Arief | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2607.20216v1

This paper investigates whether orchestrated ensembles of small language models can match or exceed single LLM performance on malware detonation report analysis. Evaluating four orchestration architectures against 11 SLMs, 3 cybersecurity models, and 6 frontier LLMs on Meta’s CyberSecEval benchmark, a hybrid system (Qwen3-4B with Foundation-Sec-8B) achieved 35.3% accuracy, exceeding the strongest cyber-specialized baseline (22.5%) and ungrounded frontier baselines (34.8%). Grounded Gemini remained the strongest single configuration at 38.2%.

Takeaway: A strong case for collaborative small-model ensembles in specialized security domains, showing that careful orchestration can approach frontier model performance at a fraction of the cost.

7. Closing the Lab-to-Store Gap: A Data-Efficient Post-Training and Experience-Driven Learning VLA Framework for Retail Humanoids

Authors: Roger Sala Sisó, Tiago Silvério, Jakob Sand, Tran Nguyen Le | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2607.20345v1

This paper presents DEED, a systems-level approach for Vision-Language-Action humanoid robots evaluated on a supermarket chip-restocking task using a Unitree G1-Edu robot and the GR00T N1.6 foundation model. The framework comprises a data-efficient post-training pipeline, experience-driven refinement via a text-based advantage prefix, and a latent-space analysis tool for distribution behavior. Results suggest that bridging the lab-to-store gap is primarily a systems integration challenge rather than an architectural one.

Takeaway: A grounded, practical demonstration that careful data design and targeted post-training—not model architecture—are the key bottlenecks for deploying humanoid robots in real-world retail environments.

8. Statevector-Referenced Geometry Survival of a Four-Qubit ZZ Quantum Kernel on IBM Quantum Hardware

Authors: Rostyslav Sipakov | Categories: quant-ph, cs.LG Link: arxiv.org/abs/2607.20377v1

This paper measures how well a four-qubit ZZ quantum kernel’s geometry survives execution on IBM quantum hardware under baseline, dynamical decoupling, and gate twirling configurations. While all configurations preserved geometry substantially (CKA 0.933-0.989), gate twirling was most faithful, though fidelity and label alignment were reversed—the most faithful configuration had the lowest kernel-target alignment. The paper emphasizes that implementation fidelity and task relevance are distinct axes that hardware quantum ML studies should report separately.

Takeaway: A careful, honest diagnostic study that disentangles hardware fidelity from task performance, providing important methodological guidance for the quantum machine learning community.

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