arXiv subjects

Dani Yogatama

Publications and source records attributed to Dani Yogatama.

At least 19 recordsLinked to original sources

arXiv2026

Rigorous Interpretation Is a Form of Evaluation

Current machine learning models are evaluated through behavioral snapshots, with benchmark accuracies, win rates and outcome-based metrics. Model explanations and evaluations, however, are fundamentally intertwined: understanding why a model produces a behavior can be as important as measuring what it produces. If we trusted interpretability, we argue that it can serve not merely as diagnostics but as a richer and more principled form of model evaluation beyond surface-level performance metrics. We explore three ways interpretability can function evaluatively: (1) fixing problems by identifying the root causes of unwanted behavior, (2) detecting subtly faulty mechanisms that invalidate model outputs, and (3) predicting potential issues before they arise by fully understanding the model's weaknesses. To fulfill its evaluative potential, we argue that interpretability methods must generate claims that are falsifiable, reproducible, and predictive -- that is, interpretability must meet scientific standards.

cs.CY↗

arXiv2026

FOL-Traces: Verified First-Order Logic Reasoning Traces at Scale

Reasoning in language models is difficult to evaluate: natural-language traces are unverifiable, symbolic datasets are too small, and most benchmarks conflate heuristics with inference. We present FOL-Traces, the first large-scale dataset of programmatically verified reasoning traces, enabling rigorous evaluation of structured logical inference. We also propose two challenging and comprehensive diagnostic tasks-masked operation prediction and step completion-that directly probe syntactic awareness and process fidelity. FOL-Traces serves as a scalable testbed for rigorously studying how models perform structured logical inference. Systematic experiments with 5 reasoning LLMs show that the dataset remains challenging: models only reach around 45.7% accuracy on masked operation prediction and around 27% on two-step completion.

cs.AI↗

arXiv2026

Evaluating Large Language Models for Fair and Reliable Organ Allocation

Medical institutions are considering the use of LLMs in high-stakes clinical decision-making, such as organ allocation. In such sensitive use cases, evaluating fairness is imperative. However, existing evaluation methods often fall short; benchmarks are too simplistic to capture real-world complexity, and accuracy-based metrics fail to address the absence of a clear ground truth. To realistically and fairly model organ allocation, specifically kidney allocation, we begin by testing the medical knowledge of LLMs to determine whether they understand the clinical factors required to make sound allocation decisions. Building on this foundation, we design two tasks: (1) Choose-One and (2) Rank-All. In Choose-One, LLMs select a single candidate from a list of potential candidates to receive a kidney. In this scenario, we assess fairness across demographics using traditional fairness metrics, such as proportional parity. In Rank-All, LLMs rank all candidates waiting for a kidney, reflecting real-world allocation processes more closely, where an organ is passed down a ranked list until allocated. Our evaluation on three LLMs reveals a divergence between fairness metrics: while exposure-based metrics suggest equitable outcomes, probability-based metrics uncover systematic preferential sorting, where specific groups were clustered in upper-ranking tiers. Furthermore, we observe that demographic preferences are highly task-dependent, showing inverted trends between Choose-One and Rank-All tasks, even when considering the topmost rank. Overall, our results indicate that current LLMs can introduce inequalities in real-world allocation scenarios, underscoring the urgent need for rigorous fairness evaluation and human oversight before their use in high-stakes decision-making.

cs.LG↗

arXiv2026

Pelican Soup Framework: A Theoretical Framework for Language Model Capabilities

In this work, we propose a simple theoretical framework, Pelican Soup, aiming to better understand how pretraining allows LLMs to (1) generalize to unseen instructions and (2) perform in-context learning, even when the verbalizers are irrelevant to the task. To this end, in our framework, we introduce the notion of "knowledge base" and "reference-sense association" and a simple formalism for natural language processing tasks. Our framework demonstrates how linguistic, psychology, and philosophy studies can inform our understanding of the language model and is connected to several other existing theoretical results. As an illustration of the usage of our framework, we derive a bound on in-context learning loss with our framework. Finally, we support our framework with empirical experiments and provide possible future research directions.

cs.CL↗

arXiv2025

Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models

Steering methods have emerged as effective and targeted tools for guiding large language models' (LLMs) behavior without modifying their parameters. Multimodal large language models (MLLMs), however, do not currently enjoy the same suite of techniques, due in part to their recency and architectural diversity. Inspired by this gap, we investigate whether MLLMs can be steered using vectors derived from their text-only LLM backbone, via sparse autoencoders (SAEs), mean shift, and linear probing. We find that text-derived steering consistently enhances multimodal accuracy across diverse MLLM architectures and visual tasks. In particular, mean shift boosts spatial relationship accuracy on CV-Bench by up to +7.3% and counting accuracy by up to +3.3%, outperforming prompting and exhibiting strong generalization to out-of-distribution datasets. These results highlight textual steering vectors as a powerful, efficient mechanism for enhancing grounding in MLLMs with minimal additional data collection and computational overhead.

cs.LG↗

arXiv2025

The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities

Modern language models can process inputs across diverse languages and modalities. We hypothesize that models acquire this capability through learning a shared representation space across heterogeneous data types (e.g., different languages and modalities), which places semantically similar inputs near one another, even if they are from different modalities/languages. We term this the semantic hub hypothesis, following the hub-and-spoke model from neuroscience (Patterson et al., 2007) which posits that semantic knowledge in the human brain is organized through a transmodal semantic "hub" which integrates information from various modality-specific "spokes" regions. We first show that model representations for semantically equivalent inputs in different languages are similar in the intermediate layers, and that this space can be interpreted using the model's dominant pretraining language via the logit lens. This tendency extends to other data types, including arithmetic expressions, code, and visual/audio inputs. Interventions in the shared representation space in one data type also predictably affect model outputs in other data types, suggesting that this shared representations space is not simply a vestigial byproduct of large-scale training on broad data, but something that is actively utilized by the model during input processing.

cs.CL↗

arXiv2025

The Rotary Position Embedding May Cause Dimension Inefficiency in Attention Heads for Long-Distance Retrieval

The Rotary Position Embedding (RoPE) is widely used in the attention heads of many large language models (LLM). It rotates dimensions in the query and the key vectors by different angles according to their positions in the input sequence. For long context modeling, the range of positions may vary a lot, and thus RoPE rotates some dimensions by a great range of angles. We hypothesize that the wide range of rotation angles may prevent LLMs from utilizing those dimensions. To validate this hypothesis, we present a controlled experiment showing that applying RoPE causes low utility of certain dimensions. Our analyses on three LLMs also indicate that these dimensions do not help LLMs do long-context question answering.

cs.CL↗

arXiv2024

Causal Interventions on Causal Paths: Mapping GPT-2's Reasoning From Syntax to Semantics

While interpretability research has shed light on some internal algorithms utilized by transformer-based LLMs, reasoning in natural language, with its deep contextuality and ambiguity, defies easy categorization. As a result, formulating clear and motivating questions for circuit analysis that rely on well-defined in-domain and out-of-domain examples required for causal interventions is challenging. Although significant work has investigated circuits for specific tasks, such as indirect object identification (IOI), deciphering natural language reasoning through circuits remains difficult due to its inherent complexity. In this work, we take initial steps to characterize causal reasoning in LLMs by analyzing clear-cut cause-and-effect sentences like "I opened an umbrella because it started raining," where causal interventions may be possible through carefully crafted scenarios using GPT-2 small. Our findings indicate that causal syntax is localized within the first 2-3 layers, while certain heads in later layers exhibit heightened sensitivity to nonsensical variations of causal sentences. This suggests that models may infer reasoning by (1) detecting syntactic cues and (2) isolating distinct heads in the final layers that focus on semantic relationships.

cs.CL↗

arXiv2024

LocateBench: Evaluating the Locating Ability of Vision Language Models

The ability to locate an object in an image according to natural language instructions is crucial for many real-world applications. In this work we propose LocateBench, a high-quality benchmark dedicated to evaluating this ability. We experiment with multiple prompting approaches, and measure the accuracy of several large vision language models. We find that even the accuracy of the strongest model, GPT-4o, lags behind human accuracy by more than 10%.

cs.CV↗

arXiv2024

DeLLMa: Decision Making Under Uncertainty with Large Language Models

The potential of large language models (LLMs) as decision support tools is increasingly being explored in fields such as business, engineering, and medicine, which often face challenging tasks of decision-making under uncertainty. In this paper, we show that directly prompting LLMs on these types of decision-making problems can yield poor results, especially as the problem complexity increases. To aid in these tasks, we propose DeLLMa (Decision-making Large Language Model assistant), a framework designed to enhance decision-making accuracy in uncertain environments. DeLLMa involves a multi-step reasoning procedure that integrates recent best practices in scaling inference-time reasoning, drawing upon principles from decision theory and utility theory, to provide an accurate and human-auditable decision-making process. We validate our procedure on multiple realistic decision-making environments, demonstrating that DeLLMa can consistently enhance the decision-making performance of leading language models, and achieve up to a 40% increase in accuracy over competing methods. Additionally, we show how performance improves when scaling compute at test time, and carry out human evaluations to benchmark components of DeLLMa.

cs.AI↗

arXiv2024

IsoBench: Benchmarking Multimodal Foundation Models on Isomorphic Representations

Current foundation models exhibit impressive capabilities when prompted either with text only or with both image and text inputs. But do their capabilities change depending on the input modality? In this work, we propose $\textbf{IsoBench}$, a benchmark dataset containing problems from four major areas: math, science, algorithms, and games. Each example is presented with multiple $\textbf{isomorphic representations}$ of inputs, such as visual, textual, and mathematical presentations. IsoBench provides fine-grained feedback to diagnose performance gaps caused by the form of the representation. Across various foundation models, we observe that on the same problem, models have a consistent preference towards textual representations. Most prominently, when evaluated on all IsoBench problems, Claude-3 Opus performs 28.7 points worse when provided with images instead of text; similarly, GPT-4 Turbo is 18.7 points worse and Gemini Pro is 14.9 points worse. Finally, we present two prompting techniques, $\textit{IsoCombination}$ and $\textit{IsoScratchPad}$, which improve model performance by considering combinations of, and translations between, different input representations.

cs.AI↗

arXiv2024

Interpretable Diffusion via Information Decomposition

Denoising diffusion models enable conditional generation and density modeling of complex relationships like images and text. However, the nature of the learned relationships is opaque making it difficult to understand precisely what relationships between words and parts of an image are captured, or to predict the effect of an intervention. We illuminate the fine-grained relationships learned by diffusion models by noticing a precise relationship between diffusion and information decomposition. Exact expressions for mutual information and conditional mutual information can be written in terms of the denoising model. Furthermore, pointwise estimates can be easily estimated as well, allowing us to ask questions about the relationships between specific images and captions. Decomposing information even further to understand which variables in a high-dimensional space carry information is a long-standing problem. For diffusion models, we show that a natural non-negative decomposition of mutual information emerges, allowing us to quantify informative relationships between words and pixels in an image. We exploit these new relations to measure the compositional understanding of diffusion models, to do unsupervised localization of objects in images, and to measure effects when selectively editing images through prompt interventions.

cs.LG↗

arXiv2024

Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models

We introduce Vibe-Eval: a new open benchmark and framework for evaluating multimodal chat models. Vibe-Eval consists of 269 visual understanding prompts, including 100 of hard difficulty, complete with gold-standard responses authored by experts. Vibe-Eval is open-ended and challenging with dual objectives: (i) vibe checking multimodal chat models for day-to-day tasks and (ii) rigorously testing and probing the capabilities of present frontier models. Notably, our hard set contains >50% questions that all frontier models answer incorrectly. We explore the nuances of designing, evaluating, and ranking models on ultra challenging prompts. We also discuss trade-offs between human and automatic evaluation, and show that automatic model evaluation using Reka Core roughly correlates to human judgment. We offer free API access for the purpose of lightweight evaluation and plan to conduct formal human evaluations for public models that perform well on the Vibe-Eval's automatic scores. We release the evaluation code and data, see https://github.com/reka-ai/reka-vibe-eval

cs.CL↗

arXiv2024

Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models

We introduce Reka Core, Flash, and Edge, a series of powerful multimodal language models trained from scratch by Reka. Reka models are able to process and reason with text, images, video, and audio inputs. This technical report discusses details of training some of these models and provides comprehensive evaluation results. We show that Reka Edge and Reka Flash are not only state-of-the-art but also outperform many much larger models, delivering outsized values for their respective compute class. Meanwhile, our most capable and largest model, Reka Core, approaches the best frontier models on both automatic evaluations and blind human evaluations. On image question answering benchmarks (e.g. MMMU, VQAv2), Core performs competitively to GPT4-V. Meanwhile, on multimodal chat, Core ranks as the second most preferred model under a blind third-party human evaluation setup, outperforming other models such as Claude 3 Opus. On text benchmarks, Core not only performs competitively to other frontier models on a set of well-established benchmarks (e.g. MMLU, GSM8K) but also outperforms GPT4-0613 on human evaluation. On video question answering (Perception-Test), Core outperforms Gemini Ultra. Models are shipped in production at http://chat.reka.ai . A showcase of non cherry picked qualitative examples can also be found at http://showcase.reka.ai .

cs.CL↗

arXiv2024

On Retrieval Augmentation and the Limitations of Language Model Training

Augmenting a language model (LM) with $k$-nearest neighbors ($k$NN) retrieval on its training data alone can decrease its perplexity, though the underlying reasons for this remain elusive. In this work, we rule out one previously posited possibility -- the "softmax bottleneck." We then create a new dataset to evaluate LM generalization ability in the setting where training data contains additional information that is not causally relevant. This task is challenging even for GPT-3.5 Turbo. We show that, for both GPT-2 and Mistral 7B, $k$NN retrieval augmentation consistently improves performance in this setting. Finally, to make $k$NN retrieval more accessible, we propose using a multi-layer perceptron model that maps datastore keys to values as a drop-in replacement for traditional retrieval. This reduces storage costs by over 25x.

cs.CL↗

arXiv2023

The Distributional Hypothesis Does Not Fully Explain the Benefits of Masked Language Model Pretraining

We analyze the masked language modeling pretraining objective function from the perspective of the distributional hypothesis. We investigate whether better sample efficiency and the better generalization capability of models pretrained with masked language modeling can be attributed to the semantic similarity encoded in the pretraining data's distributional property. Via a synthetic dataset, our analysis suggests that distributional property indeed leads to the better sample efficiency of pretrained masked language models, but does not fully explain the generalization capability. We also conduct analyses over two real-world datasets and demonstrate that the distributional property does not explain the generalization ability of pretrained natural language models either. Our results illustrate our limited understanding of model pretraining and provide future research directions.

cs.CL↗

arXiv2023

High-Modality Multimodal Transformer: Quantifying Modality & Interaction Heterogeneity for High-Modality Representation Learning

Many real-world problems are inherently multimodal, from spoken language, gestures, and paralinguistics humans use to communicate, to force, proprioception, and visual sensors on robots. While there has been an explosion of interest in multimodal learning, these methods are focused on a small set of modalities primarily in language, vision, and audio. In order to accelerate generalization towards diverse and understudied modalities, this paper studies efficient representation learning for high-modality scenarios involving a large set of diverse modalities. Since adding new models for every new modality becomes prohibitively expensive, a critical technical challenge is heterogeneity quantification: how can we measure which modalities encode similar information and interactions in order to permit parameter sharing with previous modalities? This paper proposes two new information theoretic metrics for heterogeneity quantification: (1) modality heterogeneity studies how similar 2 modalities {X1,X2} are by measuring how much information can be transferred from X1 to X2, while (2) interaction heterogeneity studies how similarly pairs of modalities {X1,X2}, {X3,X4} interact by measuring how much information can be transferred from fusing {X1,X2} to {X3,X4}. We show the importance of these 2 proposed metrics as a way to automatically prioritize the fusion of modalities that contain unique information or interactions. The result is a single model, HighMMT, that scales up to 10 modalities (text, image, audio, video, sensors, proprioception, speech, time-series, sets, and tables) and 15 tasks from 5 research areas. Not only does HighMMT outperform prior methods on the tradeoff between performance and efficiency, it also demonstrates a crucial scaling behavior: performance continues to improve with each modality added, and it transfers to entirely new modalities and tasks during fine-tuning.

cs.LG↗

arXiv2023

Questions Are All You Need to Train a Dense Passage Retriever

We introduce ART, a new corpus-level autoencoding approach for training dense retrieval models that does not require any labeled training data. Dense retrieval is a central challenge for open-domain tasks, such as Open QA, where state-of-the-art methods typically require large supervised datasets with custom hard-negative mining and denoising of positive examples. ART, in contrast, only requires access to unpaired inputs and outputs (e.g. questions and potential answer documents). It uses a new document-retrieval autoencoding scheme, where (1) an input question is used to retrieve a set of evidence documents, and (2) the documents are then used to compute the probability of reconstructing the original question. Training for retrieval based on question reconstruction enables effective unsupervised learning of both document and question encoders, which can be later incorporated into complete Open QA systems without any further finetuning. Extensive experiments demonstrate that ART obtains state-of-the-art results on multiple QA retrieval benchmarks with only generic initialization from a pre-trained language model, removing the need for labeled data and task-specific losses.

cs.CL↗