Mixture-of-Experts (MoE) models face deployment challenges due to their large parameter counts and computational demands. We explore quantization for MoE models and highlight two key insights: 1) linear blocks exhibit varying quantization sensitivity, and 2) divergent expert activation frequencies create heterogeneous computational characteristics. Based on these observations, we introduce MxMoE, a mixed-precision optimization framework for MoE models that considers both algorithmic and system perspectives. MxMoE navigates the design space defined by parameter sensitivity, expert activation dynamics, and hardware resources to derive efficient mixed-precision configurations. Additionally, MxMoE automatically generates optimized mixed-precision GroupGEMM kernels, enabling parallel execution of GEMMs with different precisions. Evaluations show that MxMoE outperforms existing methods, achieving 2.4 lower Wikitext-2 perplexity than GPTQ at 2.25-bit and delivering up to 3.4x speedup over full precision, as well as up to 29.4% speedup over uniform quantization at equivalent accuracy with 5-bit weight-activation quantization. Our code is available at https://github.com/cat538/MxMoE.

Large language models (LLMs) are rapidly pushing the limits of contemporary computing hardware. For example, training GPT-3 has been estimated to consume around 1300 MWh of electricity, and projections suggest future models may require city-scale (gigawatt) power budgets. These demands motivate exploration of computing paradigms beyond conventional von Neumann architectures. This review surveys emerging photonic hardware optimized for next-generation generative AI computing. We discuss integrated photonic neural network architectures (e.g., Mach-Zehnder interferometer meshes, lasers, wavelength-multiplexed microring resonators) that perform ultrafast matrix operations. We also examine promising alternative neuromorphic devices, including spiking neural network circuits and hybrid spintronic-photonic synapses, which combine memory and processing. The integration of two-dimensional materials (graphene, TMDCs) into silicon photonic platforms is reviewed for tunable modulators and on-chip synaptic elements. Transformer-based LLM architectures (self-attention and feed-forward layers) are analyzed in this context, identifying strategies and challenges for mapping dynamic matrix multiplications onto these novel hardware substrates. We then dissect the mechanisms of mainstream LLMs, such as ChatGPT, DeepSeek, and LLaMA, highlighting their architectural similarities and differences. We synthesize state-of-the-art components, algorithms, and integration methods, highlighting key advances and open issues in scaling such systems to mega-sized LLM models. We find that photonic computing systems could potentially surpass electronic processors by orders of magnitude in throughput and energy efficiency, but require breakthroughs in memory, especially for long-context windows and long token sequences, and in storage of ultra-large datasets.

Transformer-based models are the foundation of modern machine learning, but their execution, particularly during autoregressive decoding in large language models (LLMs), places significant pressure on memory systems due to frequent memory accesses and growing key-value (KV) caches. This creates a bottleneck in memory bandwidth, especially as context lengths increase. Processing-in-memory (PIM) architectures are a promising solution, offering high internal bandwidth and compute parallelism near memory. However, current PIM designs are primarily optimized for dense attention and struggle with the dynamic, irregular access patterns introduced by modern KV cache sparsity techniques. Consequently, they suffer from workload imbalance, reducing throughput and resource utilization. In this work, we propose STARC, a novel sparsity-optimized data mapping scheme tailored specifically for efficient LLM decoding on PIM architectures. STARC clusters KV pairs by semantic similarity and maps them to contiguous memory regions aligned with PIM bank structures. During decoding, queries retrieve relevant tokens at cluster granularity by matching against precomputed centroids, enabling selective attention and parallel processing without frequent reclustering or data movement overhead. Experiments on the HBM-PIM system show that, compared to common token-wise sparsity methods, STARC reduces attention-layer latency by 19%--31% and energy consumption by 19%--27%. Under a KV cache budget of 1024, it achieves up to 54%--74% latency reduction and 45%--67% energy reduction compared to full KV cache retrieval. Meanwhile, STARC maintains model accuracy comparable to state-of-the-art sparse attention methods, demonstrating its effectiveness in enabling efficient and hardware-friendly long-context LLM inference on PIM architectures.

Supervised and preference-based fine-tuning techniques have become popular for aligning large language models (LLMs) with user intent and correctness criteria. However, real-world training data often exhibits spurious correlations -- arising from biases, dataset artifacts, or other "shortcut" features -- that can compromise a model's performance or generalization. In this paper, we systematically evaluate three post-training algorithms -- Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and KTO (Kahneman-Tversky Optimization) -- across a diverse set of synthetic tasks and spuriousness conditions. Our tasks span mathematical reasoning, constrained instruction-following, and document-grounded question answering. We vary the degree of spurious correlation (10% vs. 90%) and investigate two forms of artifacts: "Feature Ambiguity" and "Distributional Narrowness." Our results show that the models often but not always degrade under higher spuriousness. The preference-based methods (DPO/KTO) can demonstrate relative robustness in mathematical reasoning tasks. By contrast, SFT maintains stronger performance in complex, context-intensive tasks. These findings highlight that no single post-training strategy universally outperforms in all scenarios; the best choice depends on the type of target task and the nature of spurious correlations.

Retrieval-Augmented Generation (RAG) has become a popular technique for enhancing the reliability and utility of Large Language Models (LLMs) by grounding responses in external documents. Traditional RAG systems rely on Optical Character Recognition (OCR) to first process scanned documents into text. However, even state-of-the-art OCRs can introduce errors, especially in degraded or complex documents. Recent vision-language approaches, such as ColPali, propose direct visual embedding of documents, eliminating the need for OCR. This study presents a systematic comparison between a vision-based RAG system (ColPali) and more traditional OCR-based pipelines utilizing Llama 3.2 (90B) and Nougat OCR across varying document qualities. Beyond conventional retrieval accuracy metrics, we introduce a semantic answer evaluation benchmark to assess end-to-end question-answering performance. Our findings indicate that while vision-based RAG performs well on documents it has been fine-tuned on, OCR-based RAG is better able to generalize to unseen documents of varying quality. We highlight the key trade-offs between computational efficiency and semantic accuracy, offering practical guidance for RAG practitioners in selecting between OCR-dependent and vision-based document retrieval systems in production environments.

In this paper, we introduce RWKV-X, a novel hybrid architecture that combines the efficiency of RWKV for short-range modeling with a sparse attention mechanism designed to capture long-range context. Unlike previous hybrid approaches that rely on full attention layers and retain quadratic complexity, RWKV-X achieves linear-time complexity in training and constant-time complexity in inference decoding. We demonstrate that RWKV-X, when continually pretrained on 64K-token sequences, achieves near-perfect accuracy on the 64K passkey retrieval benchmark. It consistently outperforms prior RWKV-7 models on long-context benchmarks, while maintaining strong performance on short-context tasks. These results highlight RWKV-X as a scalable and efficient backbone for general-purpose language modeling, capable of decoding sequences up to 1 million tokens with stable speed and memory usage. To facilitate further research and analysis, we have made the checkpoints and the associated code publicly accessible at: https://github.com/howard-hou/RWKV-X.

Predicting enzymatic reactions is crucial for applications in biocatalysis, metabolic engineering, and drug discovery, yet it remains a complex and resource-intensive task. Large Language Models (LLMs) have recently demonstrated remarkable success in various scientific domains, e.g., through their ability to generalize knowledge, reason over complex structures, and leverage in-context learning strategies. In this study, we systematically evaluate the capability of LLMs, particularly the Llama-3.1 family (8B and 70B), across three core biochemical tasks: Enzyme Commission number prediction, forward synthesis, and retrosynthesis. We compare single-task and multitask learning strategies, employing parameter-efficient fine-tuning via LoRA adapters. Additionally, we assess performance across different data regimes to explore their adaptability in low-data settings. Our results demonstrate that fine-tuned LLMs capture biochemical knowledge, with multitask learning enhancing forward- and retrosynthesis predictions by leveraging shared enzymatic information. We also identify key limitations, for example challenges in hierarchical EC classification schemes, highlighting areas for further improvement in LLM-driven biochemical modeling.

Drug resistance presents a major challenge in cancer therapy. Single cell profiling offers insights into cellular heterogeneity, yet the application of large-scale foundation models for predicting drug response in single cell data remains underexplored. To address this, we developed scDrugMap, an integrated framework featuring both a Python command-line interface and a web server for drug response prediction. scDrugMap evaluates a wide range of foundation models, including eight single-cell models and two large language models, using a curated dataset of over 326,000 cells in the primary collection and 18,800 cells in the validation set, spanning 36 datasets and diverse tissue and cancer types. We benchmarked model performance under pooled-data and cross-data evaluation settings, employing both layer freezing and Low-Rank Adaptation (LoRA) fine-tuning strategies. In the pooled-data scenario, scFoundation achieved the best performance, with mean F1 scores of 0.971 (layer freezing) and 0.947 (fine-tuning), outperforming the lowest-performing model by over 50%. In the cross-data setting, UCE excelled post fine-tuning (mean F1: 0.774), while scGPT led in zero-shot learning (mean F1: 0.858). Overall, scDrugMap provides the first large-scale benchmark of foundation models for drug response prediction in single-cell data and serves as a user-friendly, flexible platform for advancing drug discovery and translational research.

Mutation testing is a widely recognized technique for assessing and enhancing the effectiveness of software test suites by introducing deliberate code mutations. However, its application often results in overly large test suites, as developers generate numerous tests to kill specific mutants, increasing computational overhead. This paper introduces PRIMG (Prioritization and Refinement Integrated Mutation-driven Generation), a novel framework for incremental and adaptive test case generation for Solidity smart contracts. PRIMG integrates two core components: a mutation prioritization module, which employs a machine learning model trained on mutant subsumption graphs to predict the usefulness of surviving mutants, and a test case generation module, which utilizes Large Language Models (LLMs) to generate and iteratively refine test cases to achieve syntactic and behavioral correctness. We evaluated PRIMG on real-world Solidity projects from Code4Arena to assess its effectiveness in improving mutation scores and generating high-quality test cases. The experimental results demonstrate that PRIMG significantly reduces test suite size while maintaining high mutation coverage. The prioritization module consistently outperformed random mutant selection, enabling the generation of high-impact tests with reduced computational effort. Furthermore, the refining process enhanced the correctness and utility of LLM-generated tests, addressing their inherent limitations in handling edge cases and complex program logic.

As frontier AI systems advance toward transformative capabilities, we need a parallel transformation in how we measure and evaluate these systems to ensure safety and inform governance. While benchmarks have been the primary method for estimating model capabilities, they often fail to establish true upper bounds or predict deployment behavior. This literature review consolidates the rapidly evolving field of AI safety evaluations, proposing a systematic taxonomy around three dimensions: what properties we measure, how we measure them, and how these measurements integrate into frameworks. We show how evaluations go beyond benchmarks by measuring what models can do when pushed to the limit (capabilities), the behavioral tendencies exhibited by default (propensities), and whether our safety measures remain effective even when faced with subversive adversarial AI (control). These properties are measured through behavioral techniques like scaffolding, red teaming and supervised fine-tuning, alongside internal techniques such as representation analysis and mechanistic interpretability. We provide deeper explanations of some safety-critical capabilities like cybersecurity exploitation, deception, autonomous replication, and situational awareness, alongside concerning propensities like power-seeking and scheming. The review explores how these evaluation methods integrate into governance frameworks to translate results into concrete development decisions. We also highlight challenges to safety evaluations - proving absence of capabilities, potential model sandbagging, and incentives for "safetywashing" - while identifying promising research directions. By synthesizing scattered resources, this literature review aims to provide a central reference point for understanding AI safety evaluations.