Large language models (LLMs) often fail to scale their performance on long-context tasks performance in line with the context lengths they support. This gap is commonly attributed to retrieval failures -- the models' inability to identify relevant information in the long inputs. Accordingly, recent efforts often focus on evaluating and improving LLMs' retrieval performance: if retrieval is perfect, a model should, in principle, perform just as well on a long input as it does on a short one -- or should it? This paper presents findings that the answer to this question may be negative. Our systematic experiments across 5 open- and closed-source LLMs on math, question answering, and coding tasks reveal that, even when models can perfectly retrieve all relevant information, their performance still degrades substantially (13.9%--85%) as input length increases but remains well within the models' claimed lengths. This failure occurs even when the irrelevant tokens are replaced with minimally distracting whitespace, and, more surprisingly, when they are all masked and the models are forced to attend only to the relevant tokens. A similar performance drop is observed when all relevant evidence is placed immediately before the question. Our findings reveal a previously-unrealized limitation: the sheer length of the input alone can hurt LLM performance, independent of retrieval quality and without any distraction. They motivate our simple, model-agnostic mitigation strategy that transforms a long-context task into a short-context one by prompting the model to recite the retrieved evidence before attempting to solve the problem. On RULER, we observe a consistent improvement of GPT-4o up to 4% on an already strong baseline.

Quantizing the key-value (KV) cache is a promising strategy for improving the inference efficiency of large language models (LLMs). However, aggressive quantization to very low precision (e.g., 2 bits) introduces significant errors in the stored key and value tensors, which propagate through the dot-product attention mechanism and ultimately degrade generation quality. To address this, we propose KVLinC, a framework to mitigate attention errors introduced by KV cache quantization in the extreme low-precision regime. KVLinC combines a Hadamard rotation, which reduces quantization error in values, with lightweight linear correction adapters that explicitly compensate for errors introduced by quantized keys. Across extensive evaluations on the LLaMA, Qwen2.5, and Qwen3 model families, KVLinC consistently matches or surpasses strong baselines while achieving higher KV-cache compression. Furthermore, we implement a custom attention kernel that results in upto 2.55x faster inference compared to Flash Attention baseline, enabling efficient long-context LLM inference.

Transformers have dominated sequence processing tasks for the past seven years -- most notably language modeling. However, the inherent quadratic complexity of their attention mechanism remains a significant bottleneck as context length increases. This paper surveys recent efforts to overcome this bottleneck, including advances in (sub-quadratic) attention variants, recurrent neural networks, state space models, and hybrid architectures. We critically analyze these approaches in terms of compute and memory complexity, benchmark results, and fundamental limitations to assess whether the dominance of pure-attention transformers may soon be challenged.

Adapting Foundation Models to new domains with limited training data is challenging and computationally expensive. While prior work has demonstrated the effectiveness of using domain-specific exemplars as in-context demonstrations, we investigate whether representing exemplars purely as text is the most efficient, effective, and stable approach. We explore an alternative: representing exemplars as soft prompts with an exemplar order invariant model architecture. To this end, we introduce Multi-Head Attention Retrieval-Augmented Generation (MHA-RAG), a framework with the number of attention heads serving as a simple hyperparameter to control soft prompt-generation across different tasks. Across multiple question-answering benchmarks and model scales, MHA-RAG achieves a 20-point performance gain over standard RAG, while cutting inference costs by a factor of 10X GFLOPs-delivering both higher accuracy and greater efficiency, invariant to exemplar order.

Training large models with distributed data parallelism (DDP) requires frequent communication of gradients across workers, which can saturate bandwidth. Infrequent communication strategies (e.g., Local SGD) reduce this overhead but, when applied to adaptive optimizers, often suffer a performance gap relative to fully synchronous DDP. We trace this gap to a time-scale mismatch: the optimizer's fast-moving momentum, tuned for frequent updates, decays too quickly to smooth gradients over long intervals, leading to noise-dominated optimization. To address this, we propose MT-DAO, a family of optimizers that employs multiple slow- and fast-moving first momenta or the gradient to track update dynamics across different time scales, for which we provide the first convergence guarantees. Empirically, for language-model pre-training, this eliminates the performance gap with DDP, outperforming infrequent-communication baselines in perplexity and reducing iso-token wall-clock time by 6-27% on Ethernet interconnects. At the 720M scale, MT-DAO reaches a target perplexity in 24% fewer steps and 35% less time than the single-momentum DDP baseline. MT-DAO enables effective cross-datacenter training and training over wide geographic areas.

Direct Preference Optimization (DPO) has emerged as a simple and effective method for aligning large language models. However, its reliance on a fixed temperature parameter leads to suboptimal training on diverse preference data, causing overfitting on easy examples and under-learning from informative ones. Recent methods have emerged to counter this. While IPO addresses general overfitting, its uniform regularization can be overly conservative. The more targeted approach of $β$-DPO suffers from its own limitations: its batch-level adaptation applies a single, compromised temperature to mixed-margin pairs, its linear update rule can produce unstable negative $β$ values, and its filtering mechanism discards potentially useful training signals. In this work, we introduce Margin-Adaptive Direct Preference Optimization (MADPO), a method that provides a stable, data-preserving, and instance-level solution. MADPO employs a practical two-step approach: it first trains a reward model to estimate preference margins and then uses these margins to apply a continuous, adaptive weight to the DPO loss for each individual training sample. This re-weighting scheme creates an effective target margin that is amplified for hard pairs and dampened for easy pairs, allowing for granular control over the learning signal. We provide a comprehensive theoretical analysis, proving that MADPO has a well-behaved optimization landscape and is robust to reward model estimation errors. We validate our theory with experiments on a sentiment generation task, where MADPO consistently and significantly outperforms strong baselines across datasets of varying quality. It achieves performance gains of up to +33.3\% on High Quality data and +10.5\% on Low Quality data over the next-best method. Our results establish MADPO as a more robust and principled approach to preference alignment.

As large language models (LLMs) continue to be integrated into modern technology, there has been an increased push towards code generation applications, which also naturally extends to hardware design automation. LLM-based solutions for register transfer level (RTL) code generation for intellectual property (IP) designs have grown, especially with fine-tuned LLMs, prompt engineering, and agentic approaches becoming popular in literature. However, a gap has been exposed in these techniques, as they fail to integrate novel IPs into the model's knowledge base, subsequently resulting in poorly generated code. Additionally, as general-purpose LLMs continue to improve, fine-tuned methods on older models will not be able to compete to produce more accurate and efficient designs. Although some retrieval augmented generation (RAG) techniques exist to mitigate challenges presented in fine-tuning approaches, works tend to leverage low-quality codebases, incorporate computationally expensive fine-tuning in the frameworks, or do not use RAG directly in the RTL generation step. In this work, we introduce DeepV: a model-agnostic RAG framework to generate RTL designs by enhancing context through a large, high-quality dataset without any RTL-specific training. Our framework benefits the latest commercial LLM, OpenAI's GPT-5, with a near 17% increase in performance on the VerilogEval benchmark. We host DeepV for use by the community in a Hugging Face (HF) Space: https://huggingface.co/spaces/FICS-LLM/DeepV.

With the increasing adoption of large language models (LLMs), ensuring the safety of LLM systems has become a pressing concern. External LLM-based guardrail models have emerged as a popular solution to screen unsafe inputs and outputs, but they are themselves fine-tuned or prompt-engineered LLMs that are vulnerable to data distribution shifts. In this paper, taking Retrieval Augmentation Generation (RAG) as a case study, we investigated how robust LLM-based guardrails are against additional information embedded in the context. Through a systematic evaluation of 3 Llama Guards and 2 GPT -oss models, we confirmed that inserting benign documents into the guardrail context alters the judgments of input and output guardrails in around 11% and 8% of cases, making them unreliable. We separately analyzed the effect of each component in the augmented context: retrieved documents, user query, and LLM-generated response. The two mitigation methods we tested only bring minor improvements. These results expose a context-robustness gap in current guardrails and motivate training and evaluation protocols that are robust to retrieval and query composition. Large language models (LLMs) have rapidly become a central component of modern AI systems, powering applications from conversational assistants to code generation (Brown et al., 2020; Jiang et al., 2024). Their ability to generalize across domains and tasks has made them widely adopted in real-world deployments (Hadi et al., 2023). However, the same flexibility that enables their success also raises serious concerns about safety. LLMs are known to occasionally produce harmful, biased, or otherwise unsafe outputs, which poses significant risks when these models are used by millions of end users (Bai et al., 2022b; Ganguli et al., 2022; Gallegos et al., 2024; She et al., 2025; Guo et al., 2025). To mitigate such risks, the research community and industry have invested heavily in methods for aligning LLMs with human safety preferences.

As Large Language Models (LLMs) gain stronger multilingual capabilities, their ability to handle culturally diverse entities becomes crucial. Prior work has shown that LLMs often favor Western-associated entities in Arabic, raising concerns about cultural fairness. Due to the lack of multilingual benchmarks, it remains unclear if such biases also manifest in different non-Western languages. In this paper, we introduce Camellia, a benchmark for measuring entity-centric cultural biases in nine Asian languages spanning six distinct Asian cultures. Camellia includes 19,530 entities manually annotated for association with the specific Asian or Western culture, as well as 2,173 naturally occurring masked contexts for entities derived from social media posts. Using Camellia, we evaluate cultural biases in four recent multilingual LLM families across various tasks such as cultural context adaptation, sentiment association, and entity extractive QA. Our analyses show a struggle by LLMs at cultural adaptation in all Asian languages, with performance differing across models developed in regions with varying access to culturally-relevant data. We further observe that different LLM families hold their distinct biases, differing in how they associate cultures with particular sentiments. Lastly, we find that LLMs struggle with context understanding in Asian languages, creating performance gaps between cultures in entity extraction. Large Language Models (LLMs) have rapidly integrated into modern technology, serving users from diverse cultures (Adilazuarda et al., 2024). Among the vast range of text they process, LLMs frequently encounter entities such as people's names, locations, or food dishes, which are pervasive in text corpora (Wolfe & Caliskan, 2021; Pawar et al., 2025a) and often appear in user prompts (Li et al., 2024a; Wang et al., 2025). Importantly, entities carry cultural associations, making it essential for LLMs to handle culturally diverse entities fairly.

Large language models (LLMs) such as ChatGPT have evolved into powerful and ubiquitous tools. Fine-tuning on small datasets allows LLMs to acquire specialized skills for specific tasks efficiently. Although LLMs provide great utility in both general and task-specific use cases, they are limited by two security-related concerns. First, traditional LLM hardware requirements make them infeasible to run locally on consumer-grade devices. A remote network connection with the LLM provider's server is usually required, making the system vulnerable to network attacks. Second, fine-tuning an LLM for a sensitive task may involve sensitive data. Non-private fine-tuning algorithms produce models vulnerable to training data reproduction attacks. Our work addresses these security concerns by enhancing differentially private optimization algorithms and applying them to fine-tune localizable language models. We introduce adaptable gradient clipping along with other engineering enhancements to the standard DP-Adam optimizer to create DP-Adam-AC. We use our optimizer to fine-tune examples of two localizable LLM designs, small language model (Qwen2.5-0.5B) and 1.58 bit quantization (Bitnet-b1.58-2B). We demonstrate promising improvements in loss through experimentation with two synthetic datasets.