Scientific research demands sophisticated reasoning over multimodal data, a challenge especially prevalent in biology. Despite recent advances in multimodal large language models (MLLMs) for AI-assisted research, existing multimodal reasoning benchmarks only target up to college-level difficulty, while research-level benchmarks emphasize lower-level perception, falling short of the complex multimodal reasoning needed for scientific discovery. To bridge this gap, we introduce MicroVQA, a visual-question answering (VQA) benchmark designed to assess three reasoning capabilities vital in research workflows: expert image understanding, hypothesis generation, and experiment proposal. MicroVQA consists of 1,042 multiple-choice questions (MCQs) curated by biology experts across diverse microscopy modalities, ensuring VQA samples represent real scientific practice. In constructing the benchmark, we find that standard MCQ generation methods induce language shortcuts, motivating a new two-stage pipeline: an optimized LLM prompt structures question-answer pairs into MCQs; then, an agent-based `RefineBot' updates them to remove shortcuts. Benchmarking on state-of-the-art MLLMs reveal a peak performance of 53\%; models with smaller LLMs only slightly underperform top models, suggesting that language-based reasoning is less challenging than multimodal reasoning; and tuning with scientific articles enhances performance. Expert analysis of chain-of-thought responses shows that perception errors are the most frequent, followed by knowledge errors and then overgeneralization errors. These insights highlight the challenges in multimodal scientific reasoning, showing MicroVQA is a valuable resource advancing AI-driven biomedical research. MicroVQA is available at https://huggingface.co/datasets/jmhb/microvqa, and project page at https://jmhb0.github.io/microvqa.

Ensuring transparency in educational assessment is increasingly critical, particularly post-pandemic, as demand grows for fairer and more reliable evaluation methods. Comparative Judgement (CJ) offers a promising alternative to traditional assessments, yet concerns remain about its perceived opacity. This paper examines how Bayesian Comparative Judgement (BCJ) enhances transparency by integrating prior information into the judgement process, providing a structured, data-driven approach that improves interpretability and accountability. BCJ assigns probabilities to judgement outcomes, offering quantifiable measures of uncertainty and deeper insights into decision confidence. By systematically tracking how prior data and successive judgements inform final rankings, BCJ clarifies the assessment process and helps identify assessor disagreements. Multi-criteria BCJ extends this by evaluating multiple learning outcomes (LOs) independently, preserving the richness of CJ while producing transparent, granular rankings aligned with specific assessment goals. It also enables a holistic ranking derived from individual LOs, ensuring comprehensive evaluations without compromising detailed feedback. Using a real higher education dataset with professional markers in the UK, we demonstrate BCJ's quantitative rigour and ability to clarify ranking rationales. Through qualitative analysis and discussions with experienced CJ practitioners, we explore its effectiveness in contexts where transparency is crucial, such as high-stakes national assessments. We highlight the benefits and limitations of BCJ, offering insights into its real-world application across various educational settings.

Strategies for orchestrating the interactions between multiple agents, both human and artificial, can wildly overestimate performance and underestimate the cost of orchestration. We design a framework to orchestrate agents under realistic conditions, such as inference costs or availability constraints. We show theoretically that orchestration is only effective if there are performance or cost differentials between agents. We then empirically demonstrate how orchestration between multiple agents can be helpful for selecting agents in a simulated environment, picking a learning strategy in the infamous Rogers' Paradox from social science, and outsourcing tasks to other agents during a question-answer task in a user study.

Often, the needs and visual abilities differ between the annotator group and the end user group. Generating detailed diagram descriptions for blind and low-vision (BLV) users is one such challenging domain. Sighted annotators could describe visuals with ease, but existing studies have shown that direct generations by them are costly, bias-prone, and somewhat lacking by BLV standards. In this study, we ask sighted individuals to assess -- rather than produce -- diagram descriptions generated by vision-language models (VLM) that have been guided with latent supervision via a multi-pass inference. The sighted assessments prove effective and useful to professional educators who are themselves BLV and teach visually impaired learners. We release Sightation, a collection of diagram description datasets spanning 5k diagrams and 137k samples for completion, preference, retrieval, question answering, and reasoning training purposes and demonstrate their fine-tuning potential in various downstream tasks.

Confusing or otherwise unhelpful learner feedback creates or perpetuates erroneous beliefs that the teacher and learner have of each other, thereby increasing the cognitive burden placed upon the human teacher. For example, the robot's feedback might cause the human to misunderstand what the learner knows about the learning objective or how the learner learns. At the same time -- and in addition to the learning objective -- the learner might misunderstand how the teacher perceives the learner's task knowledge and learning processes. To ease the teaching burden, the learner should provide feedback that accounts for these misunderstandings and elicits efficient teaching from the human. This work endows an AI learner with a Second-order Theory of Mind that models perceived rationality as a source for the erroneous beliefs a teacher and learner may have of one another. It also explores how a learner can ease the teaching burden and improve teacher efficacy if it selects feedback which accounts for its model of the teacher's beliefs about the learner and its learning objective.

In this work, we introduce BurTorch, a compact high-performance framework designed to optimize Deep Learning (DL) training on single-node workstations through an exceptionally efficient CPU-based backpropagation (Rumelhart et al., 1986; Linnainmaa, 1970) implementation. Although modern DL frameworks rely on compilerlike optimizations internally, BurTorch takes a different path. It adopts a minimalist design and demonstrates that, in these circumstances, classical compiled programming languages can play a significant role in DL research. By eliminating the overhead of large frameworks and making efficient implementation choices, BurTorch achieves orders-of-magnitude improvements in performance and memory efficiency when computing $\nabla f(x)$ on a CPU. BurTorch features a compact codebase designed to achieve two key goals simultaneously. First, it provides a user experience similar to script-based programming environments. Second, it dramatically minimizes runtime overheads. In large DL frameworks, the primary source of memory overhead for relatively small computation graphs $f(x)$ is due to feature-heavy implementations. We benchmarked BurTorch against widely used DL frameworks in their execution modes: JAX (Bradbury et al., 2018), PyTorch (Paszke et al., 2019), TensorFlow (Abadi et al., 2016); and several standalone libraries: Autograd (Maclaurin et al., 2015), Micrograd (Karpathy, 2020), Apple MLX (Hannun et al., 2023). For small compute graphs, BurTorch outperforms best-practice solutions by up to $\times 2000$ in runtime and reduces memory consumption by up to $\times 3500$. For a miniaturized GPT-3 model (Brown et al., 2020), BurTorch achieves up to a $\times 20$ speedup and reduces memory up to $\times 80$ compared to PyTorch.

Prompt-tuning (PT) for large language models (LLMs) can facilitate the performance on various conventional NLP tasks with significantly fewer trainable parameters. However, our investigation reveals that PT provides limited improvement and may even degrade the primitive performance of LLMs on complex reasoning tasks. Such a phenomenon suggests that soft prompts can positively impact certain instances while negatively affecting others, particularly during the later phases of reasoning. To address these challenges, We first identify an information accumulation within the soft prompts. Through detailed analysis, we demonstrate that this phenomenon is often accompanied by erroneous information flow patterns in the deeper layers of the model, which ultimately lead to incorrect reasoning outcomes. we propose a novel method called \textbf{D}ynamic \textbf{P}rompt \textbf{C}orruption (DPC) to take better advantage of soft prompts in complex reasoning tasks, which dynamically adjusts the influence of soft prompts based on their impact on the reasoning process. Specifically, DPC consists of two stages: Dynamic Trigger and Dynamic Corruption. First, Dynamic Trigger measures the impact of soft prompts, identifying whether beneficial or detrimental. Then, Dynamic Corruption mitigates the negative effects of soft prompts by selectively masking key tokens that interfere with the reasoning process. We validate the proposed approach through extensive experiments on various LLMs and reasoning tasks, including GSM8K, MATH, and AQuA. Experimental results demonstrate that DPC can consistently enhance the performance of PT, achieving 4\%-8\% accuracy gains compared to vanilla prompt tuning, highlighting the effectiveness of our approach and its potential to enhance complex reasoning in LLMs.

As large language models (LLMs) become increasingly capable, ensuring that their self-generated explanations are faithful to their internal decision-making process is critical for safety and oversight. In this work, we conduct a comprehensive counterfactual faithfulness analysis across 62 models from 8 families, encompassing both pretrained and instruction-tuned variants and significantly extending prior studies of counterfactual tests. We introduce phi-CCT, a simplified variant of the Correlational Counterfactual Test, which avoids the need for token probabilities while explaining most of the variance of the original test. Our findings reveal clear scaling trends: larger models are consistently more faithful on our metrics. However, when comparing instruction-tuned and human-imitated explanations, we find that observed differences in faithfulness can often be attributed to explanation verbosity, leading to shifts along the true-positive/false-positive Pareto frontier. While instruction-tuning and prompting can influence this trade-off, we find limited evidence that they fundamentally expand the frontier of explanatory faithfulness beyond what is achievable with pretrained models of comparable size. Our analysis highlights the nuanced relationship between instruction-tuning, verbosity, and the faithful representation of model decision processes.

Large Language Models (LLMs) often generate hallucinations, producing outputs that are contextually inaccurate or factually incorrect. We introduce HICD, a novel method designed to induce hallucinations for contrastive decoding to mitigate hallucinations. Unlike existing contrastive decoding methods, HICD selects attention heads crucial to the model's prediction as inducing heads, then induces hallucinations by dispersing attention of these inducing heads and compares the hallucinated outputs with the original outputs to obtain the final result. Our approach significantly improves performance on tasks requiring contextual faithfulness, such as context completion, reading comprehension, and question answering. It also improves factuality in tasks requiring accurate knowledge recall. We demonstrate that our inducing heads selection and attention dispersion method leads to more "contrast-effective" hallucinations for contrastive decoding, outperforming other hallucination-inducing methods. Our findings provide a promising strategy for reducing hallucinations by inducing hallucinations in a controlled manner, enhancing the performance of LLMs in a wide range of tasks.

The assumption across nearly all language model (LM) tokenization schemes is that tokens should be subwords, i.e., contained within word boundaries. While providing a seemingly reasonable inductive bias, is this common practice limiting the potential of modern LMs? Whitespace is not a reliable delimiter of meaning, as evidenced by multi-word expressions (e.g., "by the way"), crosslingual variation in the number of words needed to express a concept (e.g., "spacesuit helmet" in German is "raumanzughelm"), and languages that do not use whitespace at all (e.g., Chinese). To explore the potential of tokenization beyond subwords, we introduce a "superword" tokenizer, SuperBPE, which incorporates a simple pretokenization curriculum into the byte-pair encoding (BPE) algorithm to first learn subwords, then superwords that bridge whitespace. This brings dramatic improvements in encoding efficiency: when fixing the vocabulary size to 200k, SuperBPE encodes a fixed piece of text with up to 33% fewer tokens than BPE on average. In experiments, we pretrain 8B transformer LMs from scratch while fixing the model size, vocabulary size, and train compute, varying *only* the algorithm for learning the vocabulary. Our model trained with SuperBPE achieves an average +4.0% absolute improvement over the BPE baseline across 30 downstream tasks (including +8.2% on MMLU), while simultaneously requiring 27% less compute at inference time. In analysis, we find that SuperBPE results in segmentations of text that are more uniform in per-token difficulty. Qualitatively, this may be because SuperBPE tokens often capture common multi-word expressions that function semantically as a single unit. SuperBPE is a straightforward, local modification to tokenization that improves both encoding efficiency and downstream performance, yielding better language models overall.