Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2411.19360

DENIAHL: In-Context Features Influence LLM Needle-In-A-Haystack Abilities

Hui Dai·Dan Pechi·Xinyi Yang·Garvit Banga·Raghav Mantri
View original source ↗Read source document ↗Export metadata ↓

Abstract

The Needle-in-a-haystack (NIAH) test is a general task used to assess language models' (LMs') abilities to recall particular information from long input context. This framework however does not provide a means of analyzing what factors, beyond context length, contribute to LMs' abilities or inabilities to separate and recall needles from their haystacks. To provide a systematic means of assessing what features contribute to LMs' NIAH capabilities, we developed a synthetic benchmark called DENIAHL (Data-oriented Evaluation of NIAH for LLM's). Our work expands on previous NIAH studies by ablating NIAH features beyond typical context length including data type, size, and patterns. We find stark differences between GPT-3.5 and LLaMA 2-7B's performance on DENIAHL, and drops in recall performance when features like item size are increased, and to some degree when data type is changed from numbers to letters. This has implications for increasingly large context models, demonstrating factors beyond item-number impact NIAH capabilities.

Explore related subjects

cs.CLcs.AIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Hui Dai, Dan Pechi, Xinyi Yang, Garvit Banga, Raghav Mantri. 2024-11-28. DENIAHL: In-Context Features Influence LLM Needle-In-A-Haystack Abilities. https://arxiv.org/abs/2411.19360

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models

Small models are made more capable through distillation from a larger one that shares their tokenization scheme. However, do distilled byte and token models behave similarly in terms of scaling trends as compute and data increases? To enable this comparison, we introduce two variants to efficiently convert token logits to Byte Logits: 1) approximate: Marginalize-It, and 2) exact: End-Of-Token. We then present the first large scale study of overtraining decoder-only dense transformer models varying two dimensions simultaneously: the tokenization scheme (Tokens, Bytes, Bytes w/ eot) and the training objective (Distillation vs. Cross-Entropy), sweeping layer-parameter-matched models with roughly 1 billion parameters up to 1 trillion bytes of data. Across eight benchmarks spanning three categories: Multiple Choice QA, Language Generation, and Machine Translation, we find that Token-1B models outperform byte models (End-Of-Token-1B and Bytes-1B) in the low-FLOP regime but eventually plateau; byte models start worse yet surpass Token-1B models with more compute, reaching a higher downstream task performance ceiling. Extrapolating the average top-1 error vs. validation BPB scaling laws predicts that, asymptotically, distilled End-Of-Token-1B outperforms distilled Token-1B by up to 4%. They are also far more data efficient, matching the performance of distilled Token-1B using only one-sixth of the training data. Moreover, by operating over a small vocabulary of 256 bytes instead of on the order of 100K tokens, they circumvent the need for top-k truncation during logit dumping, while also reducing logit storage costs to roughly one-fifth. Finally, our downstream performance scaling laws predict that our distilled End-Of-Token-1B models asymptotically surpass the Llama 3.2-1B, Gemma-3-1B-pt, and Gemma 2B models on averaged downstream tasks by up to 6.5%, 8.1%, and 2.1%, respectively.

cs.CL↗
arXiv2026

A Short Survey of Viewing Large Language Models in Legal Aspect

Large language models (LLMs) have transformed many fields, including natural language processing, computer vision, and reinforcement learning. These models have also made a significant impact in the field of law, where they are being increasingly utilized to automate various legal tasks, such as legal judgement prediction, legal document analysis, and legal document writing. However, the integration of LLMs into the legal field has also raised several legal problems, including privacy concerns, bias, and explainability. In this survey, we explore the integration of LLMs into the field of law. We discuss the various applications of LLMs in legal tasks, examine the legal challenges that arise from their use, and explore the data resources that can be used to specialize LLMs in the legal domain. Finally, we discuss several promising directions and conclude this paper. By doing so, we hope to provide an overview of the current state of LLMs in law and highlight the potential benefits and challenges of their integration.

cs.CL↗
arXiv2026

Customized large language models can outperform Community Notes in correcting misinformation

Addressing misinformation in real-world settings is challenging: content is often multimodal; factuality judgments are nuanced and context-dependent; new events emerge rapidly across domains; corrections must be timely, trustworthy, and politically impartial; and multidimensional, multistakeholder frameworks remain lacking. Crowdsourced fact-checking systems such as Community Notes have gained broad adoption, but timely, scalable coverage remains difficult. We introduce MUSE, which augments large language models (LLMs) with trust-aware retrieval of up-to-date evidence and task-specific multimodal reasoning. Given a piece of content, MUSE identifies whether and which parts may be false or misleading and provides explanations grounded in credible references. We also develop an evaluation framework that assesses expert-rated response quality---including identification accuracy, explanation factuality, and the relevance and credibility of supporting references---as well as user perceptions. Across social media posts spanning modalities, domains, political leanings, misinformation tactics, and popularity, MUSE consistently produces high-quality responses, including for content not previously fact-checked online, and outperforms even highly rated Community Notes by 29%. It also improves participants' recognition of misinformation by 10%. Our work establishes a general methodological and evaluative framework for timely, scalable, and trustworthy correction of misinformation.

cs.CL↗