Sean O'Brien

Tianlu Wang, Ping Yu, Xiaoqing Ellen Tan et al. · berkeley, meta-ai

As large language models improve, there is increasing interest in techniques that leverage these models' capabilities to refine their own outputs. In this work, we introduce Shepherd, a language model specifically tuned to critique responses and suggest refinements, extending beyond the capabilities of an untuned model to identify diverse errors and provide suggestions to remedy them. At the core of our approach is a high quality feedback dataset, which we curate from community feedback and human annotations. Even though Shepherd is small (7B parameters), its critiques are either equivalent or preferred to those from established models including ChatGPT. Using GPT-4 for evaluation, Shepherd reaches an average win-rate of 53-87% compared to competitive alternatives. In human evaluation, Shepherd strictly outperforms other models and on average closely ties with ChatGPT.

Abhay Gupta, Philip Meng, Ece Yurtseven et al.

Detecting biases in natural language understanding (NLU) for African American Vernacular English (AAVE) is crucial to developing inclusive natural language processing (NLP) systems. To address dialect-induced performance discrepancies, we introduce AAVENUE ({AAVE} {N}atural Language {U}nderstanding {E}valuation), a benchmark for evaluating large language model (LLM) performance on NLU tasks in AAVE and Standard American English (SAE). AAVENUE builds upon and extends existing benchmarks like VALUE, replacing deterministic syntactic and morphological transformations with a more flexible methodology leveraging LLM-based translation with few-shot prompting, improving performance across our evaluation metrics when translating key tasks from the GLUE and SuperGLUE benchmarks. We compare AAVENUE and VALUE translations using five popular LLMs and a comprehensive set of metrics including fluency, BARTScore, quality, coherence, and understandability. Additionally, we recruit fluent AAVE speakers to validate our translations for authenticity. Our evaluations reveal that LLMs consistently perform better on SAE tasks than AAVE-translated versions, underscoring inherent biases and highlighting the need for more inclusive NLP models. We have open-sourced our source code on GitHub and created a website to showcase our work at https://aavenuee.github.io.

Sean O'Brien, Mike Lewis

We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to outperform LLaMA 2, GPT-3.5 and PaLM-540B on the GSM8K math word reasoning benchmark, in addition to improvements on a collection of other tasks. Analysis suggests that Contrastive Decoding improves over existing methods by preventing some abstract reasoning errors, as well as by avoiding simpler modes such as copying sections of the input during chain-of-thought. Overall, Contrastive Decoding outperforms nucleus sampling for long-form generation and greedy decoding for reasoning tasks, making it a powerful general purpose method for generating text from language models.

Dharunish Yugeswardeenoo, Kevin Zhu, Sean O'Brien

Although LLMs have the potential to transform many fields, they still underperform humans in reasoning tasks. Existing methods induce the model to produce step-by-step calculations, but this research explores the question: Does making the LLM analyze the question improve its performance? We propose a novel prompting strategy called Question Analysis Prompting (QAP), in which the model is prompted to explain the question in $n$ words before solving. The value of $n$ influences the length of response generated by the model. QAP is evaluated on GPT 3.5 Turbo and GPT 4 Turbo on arithmetic datasets GSM8K, AQuA, and SAT and commonsense dataset StrategyQA. QAP is compared with other state-of-the-art prompts including Chain-of-Thought (CoT), Plan and Solve Prompting (PS+) and Take A Deep Breath (TADB). QAP outperforms all state-of-the-art prompts on AQuA and SAT datasets on both GPT3.5 and GPT4. QAP consistently ranks among the top-2 prompts on 75\% of the tests. A key factor of QAP performance can be attributed to response length, where detailed responses are beneficial when answering harder questions, but can negatively affect easy questions.

Ryotaro Shimizu, Takashi Wada, Yu Wang et al.

Recent research on explainable recommendation generally frames the task as a standard text generation problem, and evaluates models simply based on the textual similarity between the predicted and ground-truth explanations. However, this approach fails to consider one crucial aspect of the systems: whether their outputs accurately reflect the users' (post-purchase) sentiments, i.e., whether and why they would like and/or dislike the recommended items. To shed light on this issue, we introduce new datasets and evaluation methods that focus on the users' sentiments. Specifically, we construct the datasets by explicitly extracting users' positive and negative opinions from their post-purchase reviews using an LLM, and propose to evaluate systems based on whether the generated explanations 1) align well with the users' sentiments, and 2) accurately identify both positive and negative opinions of users on the target items. We benchmark several recent models on our datasets and demonstrate that achieving strong performance on existing metrics does not ensure that the generated explanations align well with the users' sentiments. Lastly, we find that existing models can provide more sentiment-aware explanations when the users' (predicted) ratings for the target items are directly fed into the models as input. The datasets and benchmark implementation are available at: https://github.com/jchanxtarov/sent_xrec.

Kaylee Chhua, Zhoujinyi Wen, Vedant Hathalia et al.

This study addresses the racial biases in facial expression recognition (FER) systems within Large Multimodal Foundation Models (LMFMs). Despite advances in deep learning and the availability of diverse datasets, FER systems often exhibit higher error rates for individuals with darker skin tones. Existing research predominantly focuses on traditional FER models (CNNs, RNNs, ViTs), leaving a gap in understanding racial biases in LMFMs. We benchmark four leading LMFMs: GPT-4o, PaliGemma, Gemini, and CLIP to assess their performance in facial emotion detection across different racial demographics. A linear classifier trained on CLIP embeddings obtains accuracies of 95.9\% for RADIATE, 90.3\% for Tarr, and 99.5\% for Chicago Face. Furthermore, we identify that Anger is misclassified as Disgust 2.1 times more often in Black Females than White Females. This study highlights the need for fairer FER systems and establishes a foundation for developing unbiased, accurate FER technologies. Visit https://kvjvhub.github.io/FERRacialBias/ for further information regarding the biases within facial expression recognition.

Daniel Csizmadia, Andrei Codreanu, Victor Sim et al.

We present Distill CLIP (DCLIP), a fine-tuned variant of the CLIP model that enhances multimodal image-text retrieval while preserving the original model's strong zero-shot classification capabilities. CLIP models are typically constrained by fixed image resolutions and limited context, which can hinder their effectiveness in retrieval tasks that require fine-grained cross-modal understanding. DCLIP addresses these challenges through a meta teacher-student distillation framework, where a cross-modal transformer teacher is fine-tuned to produce enriched embeddings via bidirectional cross-attention between YOLO-extracted image regions and corresponding textual spans. These semantically and spatially aligned global representations guide the training of a lightweight student model using a hybrid loss that combines contrastive learning and cosine similarity objectives. Despite being trained on only ~67,500 samples curated from MSCOCO, Flickr30k, and Conceptual Captions-just a fraction of CLIP's original dataset-DCLIP significantly improves image-text retrieval metrics (Recall@K, MAP), while retaining approximately 94% of CLIP's zero-shot classification performance. These results demonstrate that DCLIP effectively mitigates the trade-off between task specialization and generalization, offering a resource-efficient, domain-adaptive, and detail-sensitive solution for advanced vision-language tasks. Code available at https://anonymous.4open.science/r/DCLIP-B772/README.md.

Sundesh Donthi, Maximilian Spencer, Om Patel et al.

For large language models (LLMs) like NLLB and GPT, translating idioms remains a challenge. Our goal is to enhance translation fidelity by improving LLM processing of idiomatic language while preserving the original linguistic style. This has a significant social impact, as it preserves cultural nuances and ensures translated texts retain their intent and emotional resonance, fostering better cross-cultural communication. Previous work has utilized knowledge bases like IdiomKB by providing the LLM with the meaning of an idiom to use in translation. Although this method yielded better results than a direct translation, it is still limited in its ability to preserve idiomatic writing style across languages. In this research, we expand upon the knowledge base to find corresponding idioms in the target language. Our research performs translations using two methods: The first method employs the SentenceTransformers model to semantically generate cosine similarity scores between the meanings of the original and target language idioms, selecting the best idiom (Cosine Similarity method). The second method uses an LLM to find a corresponding idiom in the target language for use in the translation (LLM-generated idiom method). As a baseline, we performed a direct translation without providing additional information. Human evaluations on the English -> Chinese, and Chinese -> English show the Cosine Similarity Lookup method out-performed others in all GPT4o translations. To further build upon IdiomKB, we developed a low-resource Urdu dataset containing Urdu idioms and their translations. Despite dataset limitations, the Cosine Similarity Lookup method shows promise, potentially overcoming language barriers and enabling the exploration of diverse literary works in Chinese and Urdu.(LoResLM @ COLING Preprint)

Athena Wen, Tanush Patil, Ansh Saxena et al.

In an era where AI-driven hiring is transforming recruitment practices, concerns about fairness and bias have become increasingly important. To explore these issues, we introduce a benchmark, FAIRE (Fairness Assessment In Resume Evaluation), to test for racial and gender bias in large language models (LLMs) used to evaluate resumes across different industries. We use two methods-direct scoring and ranking-to measure how model performance changes when resumes are slightly altered to reflect different racial or gender identities. Our findings reveal that while every model exhibits some degree of bias, the magnitude and direction vary considerably. This benchmark provides a clear way to examine these differences and offers valuable insights into the fairness of AI-based hiring tools. It highlights the urgent need for strategies to reduce bias in AI-driven recruitment. Our benchmark code and dataset are open-sourced at our repository: https://github.com/athenawen/FAIRE-Fairness-Assessment-In-Resume-Evaluation.git.

Rajat Rawat, Hudson McBride, Dhiyaan Nirmal et al.

As large language models (LLMs) gain traction in healthcare, concerns about their susceptibility to demographic biases are growing. We introduce {DiversityMedQA}, a novel benchmark designed to assess LLM responses to medical queries across diverse patient demographics, such as gender and ethnicity. By perturbing questions from the MedQA dataset, which comprises medical board exam questions, we created a benchmark that captures the nuanced differences in medical diagnosis across varying patient profiles. Our findings reveal notable discrepancies in model performance when tested against these demographic variations. Furthermore, to ensure the perturbations were accurate, we also propose a filtering strategy that validates each perturbation. By releasing DiversityMedQA, we provide a resource for evaluating and mitigating demographic bias in LLM medical diagnoses.

Glenn Zhang, Treasure Mayowa, Jason Fan et al.

Producing trustworthy and reliable Large Language Models (LLMs) has become increasingly important as their usage becomes more widespread. Calibration seeks to achieve this by improving the alignment between the model's confidence and the actual likelihood of its responses being correct or desirable. However, it has been observed that the internal confidence of a model, derived from token probabilities, is not well aligned with its verbalized confidence, leading to misleading results with different calibration methods. In this paper, we propose Direct Confidence Alignment (DCA), a method using Direct Preference Optimization to align an LLM's verbalized confidence with its internal confidence rather than ground-truth accuracy, enhancing model transparency and reliability by ensuring closer alignment between the two confidence measures. We evaluate DCA across multiple open-weight LLMs on a wide range of datasets. To further assess this alignment, we also introduce three new calibration error-based metrics. Our results show that DCA improves alignment metrics on certain model architectures, reducing inconsistencies in a model's confidence expression. However, we also show that it can be ineffective on others, highlighting the need for more model-aware approaches in the pursuit of more interpretable and trustworthy LLMs.

Jay Shim, Grant Kruttschnitt, Alyssa Ma et al.

Rapidly increasing model scales coupled with steering methods such as chain-of-thought prompting have led to drastic improvements in language model reasoning. At the same time, models struggle with compositional generalization and are far from human performance on many reasoning-based benchmarks. Leveraging the success of chain-of-thought prompting, and also taking inspiration from context-aware decoding (CAD), we explore input-based contrasting methods to further encourage the type of reasoning induced by chain-of-thought prompting. While work remains to stabilize these results across datasets and models, the improvements we find warrant further investigation into input-based steering methods for context-aware reasoning.

Mrinal Agarwal, Saad Rana, Theo Sundoro et al.

Deception is a fundamental challenge for multi-agent reasoning: effective systems must strategically conceal information while detecting misleading behavior in others. Yet most evaluations reduce deception to static classification, ignoring the interactive, adversarial, and longitudinal nature of real deceptive dynamics. Large language models (LLMs) can deceive convincingly but remain weak at detecting deception in peers. We present WOLF, a multi-agent social deduction benchmark based on Werewolf that enables separable measurement of deception production and detection. WOLF embeds role-grounded agents (Villager, Werewolf, Seer, Doctor) in a programmable LangGraph state machine with strict night-day cycles, debate turns, and majority voting. Every statement is a distinct analysis unit, with self-assessed honesty from speakers and peer-rated deceptiveness from others. Deception is categorized via a standardized taxonomy (omission, distortion, fabrication, misdirection), while suspicion scores are longitudinally smoothed to capture both immediate judgments and evolving trust dynamics. Structured logs preserve prompts, outputs, and state transitions for full reproducibility. Across 7,320 statements and 100 runs, Werewolves produce deceptive statements in 31% of turns, while peer detection achieves 71-73% precision with ~52% overall accuracy. Precision is higher for identifying Werewolves, though false positives occur against Villagers. Suspicion toward Werewolves rises from ~52% to over 60% across rounds, while suspicion toward Villagers and the Doctor stabilizes near 44-46%. This divergence shows that extended interaction improves recall against liars without compounding errors against truthful roles. WOLF moves deception evaluation beyond static datasets, offering a dynamic, controlled testbed for measuring deceptive and detective capacity in adversarial multi-agent interaction.

Ishneet Sukhvinder Singh, Ritvik Aggarwal, Ibrahim Allahverdiyev et al.

Retrieval-Augmented Generation (RAG) systems using large language models (LLMs) often generate inaccurate responses due to the retrieval of irrelevant or loosely related information. Existing methods, which operate at the document level, fail to effectively filter out such content. We propose LLM-driven chunk filtering, ChunkRAG, a framework that enhances RAG systems by evaluating and filtering retrieved information at the chunk level. Our approach employs semantic chunking to divide documents into coherent sections and utilizes LLM-based relevance scoring to assess each chunk's alignment with the user's query. By filtering out less pertinent chunks before the generation phase, we significantly reduce hallucinations and improve factual accuracy. Experiments show that our method outperforms existing RAG models, achieving higher accuracy on tasks requiring precise information retrieval. This advancement enhances the reliability of RAG systems, making them particularly beneficial for applications like fact-checking and multi-hop reasoning.

Olga Golovneva, Sean O'Brien, Ramakanth Pasunuru et al. · berkeley, meta-ai

With recent advancements in large language models, methods like chain-of-thought prompting to elicit reasoning chains have been shown to improve results on reasoning tasks. However, tasks that require multiple steps of reasoning still pose significant challenges to state-of-the-art models. Drawing inspiration from the beam search algorithm, we propose PathFinder, a tree-search-based reasoning path generation approach. It enhances diverse branching and multi-hop reasoning through the integration of dynamic decoding, enabled by varying sampling methods and parameters. Using constrained reasoning, PathFinder integrates novel quality constraints, pruning, and exploration methods to enhance the efficiency and the quality of generation. Moreover, it includes scoring and ranking features to improve candidate selection. Our approach outperforms competitive baselines on three complex arithmetic and commonsense reasoning tasks by 6% on average. Our model generalizes well to longer, unseen reasoning chains, reflecting similar complexities to beam search with large branching factors.

Joshua Liu, Aarav Jain, Soham Takuri et al.

Rapid improvements in large language models have unveiled a critical challenge in human-AI interaction: sycophancy. In this context, sycophancy refers to the tendency of models to excessively agree with or flatter users, often at the expense of factual accuracy. While previous studies have primarily analyzed this behavior in single-turn interactions, its persistence and evolution in multi-step conversations remain largely unexplored. We introduce TRUTH DECAY, a benchmark specifically designed to evaluate sycophancy in extended dialogues, where language models must navigate iterative user feedback, challenges, and persuasion. We prompt models to elicit four types of sycophantic biases. We then propose and test sycophancy reduction strategies, evaluating their effectiveness beyond single-step interactions.

Muna Numan Said, Aarib Zaidi, Rabia Usman et al.

The transformative potential of text-to-image (T2I) models hinges on their ability to synthesize culturally diverse, photorealistic images from textual prompts. However, these models often perpetuate cultural biases embedded within their training data, leading to systemic misrepresentations. This paper benchmarks the Component Inclusion Score (CIS), a metric designed to evaluate the fidelity of image generation across cultural contexts. Through extensive analysis involving 2,400 images, we quantify biases in terms of compositional fragility and contextual misalignment, revealing significant performance gaps between Western and non-Western cultural prompts. Our findings underscore the impact of data imbalance, attention entropy, and embedding superposition on model fairness. By benchmarking models like Stable Diffusion with CIS, we provide insights into architectural and data-centric interventions for enhancing cultural inclusivity in AI-generated imagery. This work advances the field by offering a comprehensive tool for diagnosing and mitigating biases in T2I generation, advocating for more equitable AI systems.

McNair Shah, Saleena Angeline, Adhitya Rajendra Kumar et al.

Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and ablation in the subspace's dominant direction. We find that dominant direction steering allows for near elimination of harmfulness with a low decrease in utility. Our findings advance the emerging view that concept subspaces provide a scalable lens on LLM behaviour and offer practical tools for the community to audit and harden future generations of language models.

Anthony Cui, Pranav Nandyalam, Andrew Rufail et al.

Momentum-Aided Prompt Optimization (MAPO) enhances the efficiency and efficacy of prompt optimization for Large Language Models (LLMs). Building on ProTeGi, MAPO uses positive natural language "gradients" and a momentum-based extension to refine prompts effectively. By tracking gradient history, MAPO avoids local minima and oscillations. It also utilizes beam search and an Upper Confidence Bound (UCB) algorithm for balanced candidate expansion and selection. Benchmark testing shows that MAPO achieves faster convergence time with fewer API calls and higher F1 scores than ProTeGi, proving it as a robust and scalable solution for automated prompt engineering in LLMs.

Imran Mirza, Cole Huang, Ishwara Vasista et al.

Multi-agent systems, which consist of multiple AI models interacting within a shared environment, are increasingly used for persona-based interactions. However, if not carefully designed, these systems can reinforce implicit biases in large language models (LLMs), raising concerns about fairness and equitable representation. We present MALIBU, a novel benchmark developed to assess the degree to which LLM-based multi-agent systems implicitly reinforce social biases and stereotypes. MALIBU evaluates bias in LLM-based multi-agent systems through scenario-based assessments. AI models complete tasks within predefined contexts, and their responses undergo evaluation by an LLM-based multi-agent judging system in two phases. In the first phase, judges score responses labeled with specific demographic personas (e.g., gender, race, religion) across four metrics. In the second phase, judges compare paired responses assigned to different personas, scoring them and selecting the superior response. Our study quantifies biases in LLM-generated outputs, revealing that bias mitigation may favor marginalized personas over true neutrality, emphasizing the need for nuanced detection, balanced fairness strategies, and transparent evaluation benchmarks in multi-agent systems.

Cheng-Ting Chou, George Liu, Jessica Sun et al.

Deterministically controlling the target generation language of large multilingual language models (LLMs) remains a fundamental challenge, particularly in zero-shot settings where neither explicit language prompts nor fine-tuning are available. In this work, we investigate whether sparse autoencoder (SAE) features, previously shown to correlate with interpretable model behaviors, can be leveraged to steer the generated language of LLMs during inference. Leveraging pretrained SAEs on the residual streams of Gemma-2B and Gemma-9B, we identify features whose activations differ most significantly between English and four target languages: Chinese, Japanese, Spanish, and French. By modifying just a single SAE feature at one transformer layer, we achieve controlled language shifts with up to 90\% success, as measured by FastText language classification, while preserving semantic fidelity according to LaBSE (Language-Agnostic BERT Sentence Embedding) similarity. Our analysis reveals that language steering is most effective in mid-to-late transformer layers and is amplified by specific attention heads disproportionately associated with language-sensitive SAE features. These results demonstrate the promise of sparse feature steering as a lightweight and interpretable mechanism for controllable multilingual generation.

Jason Li, Lauren Yraola, Kevin Zhu et al.

Prompting methods for language models, such as Chain-of-thought (CoT), present intuitive step-by-step processes for problem solving. These methodologies aim to equip models with a better understanding of the correct procedures for addressing a given task. Despite these advancements, CoT lacks the ability of reflection and error correction, potentially causing a model to perpetuate mistakes and errors. Therefore, inspired by the human ability for said tasks, we propose Error Reflection Prompting (ERP) to further enhance reasoning in language models. Building upon CoT, ERP is a method comprised of an incorrect answer, error recognition, and a correct answer. This process enables the model to recognize types of errors and the steps that lead to incorrect answers, allowing the model to better discern which steps to avoid and which to take. The model is able to generate the error outlines itself with automated ERP generation, allowing for error recognition and correction to be integrated into the reasoning chain and produce scalability and reliability in the process. The results demonstrate that ERP serves as a versatile supplement to conventional CoT, ultimately contributing to more robust and capable reasoning abilities along with increased interpretability in how models ultimately reach their errors.

Ishant Chintapatla, Kazuma Choji, Naaisha Agarwal et al.

Recently, many benchmarks and datasets have been developed to evaluate Vision-Language Models (VLMs) using visual question answering (VQA) pairs, and models have shown significant accuracy improvements. However, these benchmarks rarely test the model's ability to accurately complete visual entailment, for instance, accepting or refuting a hypothesis based on the image. To address this, we propose COREVQA (Crowd Observations and Reasoning Entailment), a benchmark of 5608 image and synthetically generated true/false statement pairs, with images derived from the CrowdHuman dataset, to provoke visual entailment reasoning on challenging crowded images. Our results show that even the top-performing VLMs achieve accuracy below 80%, with other models performing substantially worse (39.98%-69.95%). This significant performance gap reveals key limitations in VLMs' ability to reason over certain types of image-question pairs in crowded scenes.

Stanley Yu, Vaidehi Bulusu, Oscar Yasunaga et al.

Large Language Models (LLMs) exhibit strong conversational abilities but often generate falsehoods. Prior work suggests that the truthfulness of simple propositions can be represented as a single linear direction in a model's internal activations, but this may not fully capture its underlying geometry. In this work, we extend the concept cone framework, recently introduced for modeling refusal, to the domain of truth. We identify multi-dimensional cones that causally mediate truth-related behavior across multiple LLM families. Our results are supported by three lines of evidence: (i) causal interventions reliably flip model responses to factual statements, (ii) learned cones generalize across model architectures, and (iii) cone-based interventions preserve unrelated model behavior. These findings reveal the richer, multidirectional structure governing simple true/false propositions in LLMs and highlight concept cones as a promising tool for probing abstract behaviors.

Arjun Prasaath Anbazhagan, Parteek Kumar, Ujjwal Kaur et al.

How does textual representation of audio relate to the Large Language Model's (LLMs) learning about the audio world? This research investigates the extent to which LLMs can be prompted to generate audio, despite their primary training in textual data. We employ a three-tier approach, progressively increasing the complexity of audio generation: 1) Musical Notes, 2) Environmental Sounds, and 3) Human Speech. To bridge the gap between text and audio, we leverage code as an intermediary, prompting LLMs to generate code that, when executed, produces the desired audio output. To evaluate the quality and accuracy of the generated audio, we employ FAD and CLAP scores. Our findings reveal that while LLMs can generate basic audio features, their performance deteriorates as the complexity of the audio increases. This suggests that while LLMs possess a latent understanding of the auditory world, their ability to translate this understanding into tangible audio output remains rudimentary. Further research into techniques that can enhance the quality and diversity of LLM-generated audio can lead to an improvement in the performance of text-based LLMs in generating audio.

Shaun Baek, Shaun Esua-Mensah, Cyrus Tsui et al.

Large Language Models (LLMs) are primarily trained on high-resource natural languages, limiting their effectiveness in low-resource settings and in tasks requiring deep logical reasoning. This research introduces Rosetta-PL, a benchmark designed to evaluate LLMs' logical reasoning and generalization capabilities in a controlled environment. We construct Rosetta-PL by translating a dataset of logical propositions from Lean into a custom logical language, which is then used to fine-tune an LLM (e.g., GPT-4o). Our experiments analyze the impact of the size of the dataset and the translation methodology on the performance of the model. Our results indicate that preserving logical relationships in the translation process significantly boosts precision, with accuracy plateauing beyond roughly 20,000 training samples. These insights provide valuable guidelines for optimizing LLM training in formal reasoning tasks and improving performance in various low-resource language applications.

Abhay Gupta, Jacob Cheung, Philip Meng et al.

The diversity of human language, shaped by social, cultural, and regional influences, presents significant challenges for natural language processing (NLP) systems. Existing benchmarks often overlook intra-language variations, leaving speakers of non-standard dialects underserved. To address this gap, we introduce EnDive (English Diversity), a benchmark that evaluates five widely-used large language models (LLMs) across tasks in language understanding, algorithmic reasoning, mathematics, and logic. Our framework translates Standard American English datasets into five underrepresented dialects using few-shot prompting with verified examples from native speakers, and compare these translations against rule-based methods via fluency assessments, preference tests, and semantic similarity metrics. Human evaluations confirm high translation quality, with average scores of at least 6.02/7 for faithfulness, fluency, and formality. By filtering out near-identical translations, we create a challenging dataset that reveals significant performance disparities - models consistently underperform on dialectal inputs compared to Standard American English. EnDive thus advances dialect-aware NLP by uncovering model biases and promoting more equitable language technologies.

James Begin, Namit Agrawal, Eshan Singh et al.

LLMs have demonstrated remarkable proficiency in understanding tasks but continue to struggle with long-context comprehension, particularly with content located in the middle of extensive inputs. This limitation, known as the Lost-in-the-Middle (LITM) problem, hinders models from fully processing and utilizing information across lengthy contexts. To address this issue, we introduce pause-tuning, a technique that redistributes attention to enhance comprehension of long-context inputs. Our approach involves fine-tuning language models on datasets with artificially inserted pause tokens, which serve to segment the input into smaller, more manageable parts. We evaluate pause-tuning against alternative approaches using the Needle-in-a-Haystack benchmark, where models must retrieve information embedded within contexts of up to 128K tokens. Experimental results demonstrate significant performance gains, with the LLaMA 3.2 3B Instruct model and the LLaMA 3.1 8B Instruct model improving by 10.61% and 3.57% respectively on average, suggesting that pause-tuning successfully enhances attention redistribution and improves long-context retention. The code and data are available at https://anonymous.4open.science/r/LITM-PauseTokens-7357.

Leo Lu, Jonathan Zhang, Sean Chua et al.

Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of large language models (LLMs). While prior work focuses on improving model performance through internal reasoning strategies, little is known about the interchangeability of reasoning across different models. In this work, we explore whether a partially completed reasoning chain from one model can be reliably continued by another model, either within the same model family or across families. We achieve this by assessing the sufficiency of intermediate reasoning traces as transferable scaffolds for logical coherence and final answer accuracy. We interpret this interchangeability as a means of examining inference-time trustworthiness, probing whether reasoning remains both coherent and reliable under model substitution. Using token-level log-probability thresholds to truncate reasoning at early, mid, and late stages from our baseline models, Gemma-3-4B-IT and LLaMA-3.1-70B-Instruct, we conduct continuation experiments with Gemma-3-1B-IT and LLaMA-3.1-8B-Instruct to test intra-family and cross-family behaviors. Our evaluation pipeline leverages truncation thresholds with a Process Reward Model (PRM), providing a reproducible framework for assessing reasoning stability via model interchange. Evaluations with a PRM reveal that hybrid reasoning chains often preserve, and in some cases even improve, final accuracy and logical structure. Our findings point towards interchangeability as an emerging behavioral property of reasoning models, offering insights into new paradigms for reliable modular reasoning in collaborative AI systems.

Claire O'Brien, Jessica Seto, Dristi Roy et al.

Behavioral alignment in large language models (LLMs) is often achieved through broad fine-tuning, which can result in undesired side effects like distributional shift and low interpretability. We propose a method for alignment that identifies and updates only the neurons most responsible for a given behavior, a targeted approach that allows for fine-tuning with significantly less data. Using sparse autoencoders (SAEs) and linear probes, we isolate the 3% of MLP neurons most predictive of a target behavior, decode them into residual space, and fine-tune only those neurons using gradient masking. We demonstrate this approach on the task of reducing sycophantic behavior, where our method matches or exceeds state-of-the-art performance on four benchmarks (Syco-Bench, NLP, POLI, PHIL) using Gemma-2-2B and 9B models. Our results show that sparse, neuron-level updates offer a scalable and precise alternative to full-model fine-tuning, remaining effective even in situations when little data is available

Aayush Aluru, Myra Malik, Samarth Patankar et al.

Multi-agent systems (MAS) often achieve higher reasoning accuracy than single models, but their reliance on repeated debates across agents makes them computationally expensive. We introduce SMAGDi, a distillation framework that transfers the debate dynamics of a five-agent Llama-based MAS into a compact Socratic decomposer-solver student. SMAGDi represents debate traces as directed interaction graphs, where nodes encode intermediate reasoning steps with correctness labels and edges capture continuity and cross-agent influence. The student is trained with a composite objective combining language modeling, graph-based supervision, contrastive reasoning, and embedding alignment to preserve both fluency and structured reasoning. On StrategyQA and MMLU, SMAGDi compresses a 40B multi-agent system into a 6B student while retaining 88% of its accuracy, substantially outperforming prior distillation methods such as MAGDi, standard KD, and fine-tuned baselines. These results highlight that explicitly modeling interaction graphs and Socratic decomposition enable small models to inherit the accuracy benefits of multi-agent debate while remaining efficient enough for real-world deployment.

Adhyayan Veer Singh, Aaron Shen, Brian Law et al.

Correctness alone is insufficient: LLM-generated programs frequently satisfy unit tests while violating contest time or memory budgets. We present SwiftSolve, a complexity-aware multi-agent system for competitive programming that couples algorithmic planning with empirical profiling and complexity-guided repair. We frame competitive programming as a software environment where specialized agents act as programmers, each assuming roles such as planning, coding, profiling, and complexity analysis. A Planner proposes an algorithmic sketch; a deterministic Static Pruner filters high-risk plans; a Coder emits ISO C++17; a Profiler compiles and executes candidates on a fixed input-size schedule to record wall time and peak memory; and a Complexity Analyst fits log-log growth (s, R2) with an LLM fallback to assign a complexity class and dispatch targeted patches to either the Planner or Coder. Agents communicate via typed, versioned JSON; a controller enforces iteration caps and diminishing returns stopping. Evaluated on 26 problems (16 BigO, 10 Codeforces Div. 2) in a POSIX sandbox (2 s / 256-512 MB), SwiftSolve attains pass@1 = 61.54% (16/26) on the first attempt and Solved@<=3 = 80.77% with marginal latency change (mean 11.96 s to 12.66 s per attempt). Aggregate run-level success is 73.08% at 12.40 s mean. Failures are predominantly resource-bound, indicating inefficiency rather than logic errors. Against Claude Opus 4, SwiftSolve improves run-level success (73.1% vs 52.6%) at approximately 2x runtime overhead (12.4 s vs 6.8 s). Beyond correctness (pass@k), we report efficiency metrics (eff@k for runtime and memory, incidence of TLE or MLE, and complexity fit accuracy on BigO), demonstrating that profiling and complexity-guided replanning reduce inefficiency while preserving accuracy.

Haziq Mohammad Khalid, Athikash Jeyaganthan, Timothy Do et al.

Large Language Models (LLMs) suffer significant performance degradation in multi-turn conversations when information is presented incrementally. Given that multi-turn conversations characterize everyday interactions with LLMs, this degradation poses a severe challenge to real world usability. We hypothesize that abrupt increases in model uncertainty signal misalignment in multi-turn LLM interactions, and we exploit this insight to dynamically realign conversational context. We introduce ERGO (Entropy-guided Resetting for Generation Optimization), which continuously quantifies internal uncertainty via Shannon entropy over next token distributions and triggers adaptive prompt consolidation when a sharp spike in entropy is detected. By treating uncertainty as a first class signal rather than a nuisance to eliminate, ERGO embraces variability in language and modeling, representing and responding to uncertainty. In multi-turn tasks with incrementally revealed instructions, ERGO yields a 56.6% average performance gain over standard baselines, increases aptitude (peak performance capability) by 24.7%, and decreases unreliability (variability in performance) by 35.3%, demonstrating that uncertainty aware interventions can improve both accuracy and reliability in conversational AI.

Duy Le, Kent Ziti, Evan Girard-Sun et al.

Language models are increasingly tested on multilingual creativity, demanding culturally grounded, abstract generations. Standard prompting methods often produce repetitive or shallow outputs. We introduce Adaptive Originality Filtering (AOF), a prompting strategy that enforces novelty and cultural fidelity via semantic rejection. To assess quality, we propose RiddleScore, a metric combining novelty, diversity, fluency, and answer alignment. AOF improves Distinct-2 (0.915 in Japanese), reduces Self-BLEU (0.177), and raises RiddleScore (up to +57.1% in Arabic). Human evaluations confirm fluency, creativity, and cultural fit gains. However, improvements vary: Arabic shows greater RiddleScore gains than Distinct-2; Japanese sees similar changes. Though focused on riddles, our method may apply to broader creative tasks. Overall, semantic filtering with composite evaluation offers a lightweight path to culturally rich generation without fine-tuning.

Advey Nandan, Cheng-Ting Chou, Amrit Kurakula et al.

We investigate the phenomenon of neuron universality in independently trained GPT-2 Small models, examining these universal neurons-neurons with consistently correlated activations across models-emerge and evolve throughout training. By analyzing five GPT-2 models at five checkpoints, we identify universal neurons through pairwise correlation analysis of activations over a dataset of 5 million tokens. Ablation experiments reveal significant functional impacts of universal neurons on model predictions, measured via cross entropy loss. Additionally, we quantify neuron persistence, demonstrating high stability of universal neurons across training checkpoints, particularly in early and deeper layers. These findings suggest stable and universal representational structures emerge during language model training.

Daniel Son, Sanjana Rathore, Andrew Rufail et al.

We investigate feature universality in Gemma-2 language models (Gemma-2-2B and Gemma-2-9B), asking whether models with a four-fold difference in scale still converge on comparable internal concepts. Using the Sparse Autoencoder (SAE) dictionary-learning pipeline, we utilize SAEs on each model's residual-stream activations, align the resulting monosemantic features via activation correlation, and compare the matched feature spaces with SVCCA and RSA. Middle layers yield the strongest overlap, while early and late layers show far less similarity. Preliminary experiments extend the analysis from single tokens to multi-token subspaces, showing that semantically similar subspaces interact similarly with language models. These results strengthen the case that large language models carve the world into broadly similar, interpretable features despite size differences, reinforcing universality as a foundation for cross-model interpretability.

Lang Xiong, Raina Gao, Alyssa Jeong et al.

Sarcasm is a form of humor where expressions convey meanings opposite to their literal interpretations. Classifying and generating sarcasm using large language models is vital for interpreting human communication. Sarcasm poses challenges for computational models, due to its nuanced nature. We introduce Sarc7, a benchmark that classifies 7 types of sarcasm: self-deprecating, brooding, deadpan, polite, obnoxious, raging, and manic by annotating entries of the MUStARD dataset. Classification was evaluated using zero-shot, few-shot, chain-of-thought (CoT), and a novel emotion-based prompting technique. We propose an emotion-based generation method developed by identifying key components of sarcasm-incongruity, shock value, and context dependency. Our classification experiments show that Gemini 2.5, using emotion-based prompting, outperforms other setups with an F1 score of 0.3664. Human evaluators preferred our emotion-based prompting, with 38.46% more successful generations than zero-shot prompting.

Timothy Do, Pranav Saran, Harshita Poojary et al.

In this paper, we address the persistent challenges that figurative language expressions pose for natural language processing (NLP) systems, particularly in low-resource languages such as Konkani. We present a hybrid model that integrates a pre-trained Multilingual BERT (mBERT) with a bidirectional LSTM and a linear classifier. This architecture is fine-tuned on a newly introduced annotated dataset for metaphor classification, developed as part of this work. To improve the model's efficiency, we implement a gradient-based attention head pruning strategy. For metaphor classification, the pruned model achieves an accuracy of 78%. We also applied our pruning approach to expand on an existing idiom classification task, achieving 83% accuracy. These results demonstrate the effectiveness of attention head pruning for building efficient NLP tools in underrepresented languages.

Abhay Gupta, Michael Lu, Kevin Zhu et al.

Current large language models (LLMs) struggle to answer questions that span tens of thousands of tokens, especially when multi-hop reasoning is involved. While prior benchmarks explore long-context comprehension or multi-hop reasoning in isolation, none jointly vary context length and reasoning depth in natural narrative settings. We introduce NovelHopQA, the first benchmark to evaluate 1-4 hop QA over 64k-128k-token excerpts from 83 full-length public-domain novels. A keyword-guided pipeline builds hop-separated chains grounded in coherent storylines. We evaluate seven state-of-the-art models and apply oracle-context filtering to ensure all questions are genuinely answerable. Human annotators validate both alignment and hop depth. We additionally present retrieval-augmented generation (RAG) evaluations to test model performance when only selected passages are provided instead of the full context. We noticed consistent accuracy drops with increased hops and context length increase, even for frontier models-revealing that sheer scale does not guarantee robust reasoning. Failure-mode analysis highlights common breakdowns such as missed final-hop integration and long-range drift. NovelHopQA offers a controlled diagnostic setting to test multi-hop reasoning at scale. All code and datasets are available at https://novelhopqa.github.io.

Jesus Alvarez C, Daua D. Karajeanes, Ashley Celeste Prado et al.

The digital exclusion of endangered languages remains a critical challenge in NLP, limiting both linguistic research and revitalization efforts. This study introduces the first computational investigation of Comanche, an Uto-Aztecan language on the verge of extinction, demonstrating how minimal-cost, community-informed NLP interventions can support language preservation. We present a manually curated dataset of 412 phrases, a synthetic data generation pipeline, and an empirical evaluation of GPT-4o and GPT-4o-mini for language identification. Our experiments reveal that while LLMs struggle with Comanche in zero-shot settings, few-shot prompting significantly improves performance, achieving near-perfect accuracy with just five examples. Our findings highlight the potential of targeted NLP methodologies in low-resource contexts and emphasize that visibility is the first step toward inclusion. By establishing a foundation for Comanche in NLP, we advocate for computational approaches that prioritize accessibility, cultural sensitivity, and community engagement.

Andrew Rufail, Daniel Kim, Sean O'Brien et al.

We introduce CLEAR (Contrasting Textual Feedback with Experts and Amateurs for Reasoning), a novel approach to language model reasoning that leverages the strengths of a larger (expert) model and smaller (amateur) model. The expert and amateur models each provide feedback on a model's initial output and are contrasted with each other into refined feedback. This feedback is subsequently applied to iteratively improve CLEAR's responses. Our experiments demonstrate that CLEAR outperforms state-of-the-art methods in several challenging reasoning tasks, including story outline improvement (up to 19.6% relative increase in interestingness), constrained generation (up to 18.5% increase in coverage), mathematical reasoning (up to 6.7% improvement in accuracy) and mitigation of toxicity (decrease of up to 22% in toxicity).