Abstract:Large Language Models (LLMs) have demonstrated remarkable text generation capabilities, and recent advances in training paradigms have led to breakthroughs in their reasoning performance. In this work, we investigate how the reasoning effort of such models scales with problem complexity. We use the infinitely scalable Tents puzzle, which has a known linear-time solution, to analyze this scaling behavior. Our results show that reasoning effort scales with problem size, but only up to a critical problem complexity. Beyond this threshold, the reasoning effort does not continue to increase, and may even decrease. This observation highlights a critical limitation in the logical coherence of current LLMs as problem complexity increases, and underscores the need for strategies to improve reasoning scalability. Furthermore, our results reveal significant performance differences between current state-of-the-art reasoning models when faced with increasingly complex logical puzzles.
Abstract:We challenge the prevailing assumption that LLMs must rely fully on sub-word tokens for high-quality text generation. To this end, we propose the "Generative Pretrained Thoughtformer" (GPTHF), a hierarchical transformer language model capable of text generation by compressing text into sentence embeddings and employing a sentence attention mechanism. GPTHF retains GPT's architecture, modifying only token interactions via dynamic sparse attention masks. Our experiments show that GPTHF achieves an up to an order of magnitude improvement in FLOPs efficiency and a threefold increase in runtime speed compared to equally-sized GPT models in the low-size regime. This is achieved through a unique generation method that caches and reuses sentence embeddings, allowing significant portions of the input to bypass large parts of the network.
Abstract:This work presents a first evaluation of two state-of-the-art Large Reasoning Models (LRMs), OpenAI's o3-mini and DeepSeek R1, on analogical reasoning, focusing on well-established nonverbal human IQ tests based on Raven's progressive matrices. We benchmark with the I-RAVEN dataset and its more difficult extension, I-RAVEN-X, which tests the ability to generalize to longer reasoning rules and ranges of the attribute values. To assess the influence of visual uncertainties on these nonverbal analogical reasoning tests, we extend the I-RAVEN-X dataset, which otherwise assumes an oracle perception. We adopt a two-fold strategy to simulate this imperfect visual perception: 1) we introduce confounding attributes which, being sampled at random, do not contribute to the prediction of the correct answer of the puzzles and 2) smoothen the distributions of the input attributes' values. We observe a sharp decline in OpenAI's o3-mini task accuracy, dropping from 86.6% on the original I-RAVEN to just 17.0% -- approaching random chance -- on the more challenging I-RAVEN-X, which increases input length and range and emulates perceptual uncertainty. This drop occurred despite spending 3.4x more reasoning tokens. A similar trend is also observed for DeepSeek R1: from 80.6% to 23.2%. On the other hand, a neuro-symbolic probabilistic abductive model, ARLC, that achieves state-of-the-art performances on I-RAVEN, can robustly reason under all these out-of-distribution tests, maintaining strong accuracy with only a modest reduction from 98.6% to 88.0%. Our code is available at https://github.com/IBM/raven-large-language-models.
Abstract:While neural vocoders have made significant progress in high-fidelity speech synthesis, their application on polyphonic music has remained underexplored. In this work, we propose DisCoder, a neural vocoder that leverages a generative adversarial encoder-decoder architecture informed by a neural audio codec to reconstruct high-fidelity 44.1 kHz audio from mel spectrograms. Our approach first transforms the mel spectrogram into a lower-dimensional representation aligned with the Descript Audio Codec (DAC) latent space before reconstructing it to an audio signal using a fine-tuned DAC decoder. DisCoder achieves state-of-the-art performance in music synthesis on several objective metrics and in a MUSHRA listening study. Our approach also shows competitive performance in speech synthesis, highlighting its potential as a universal vocoder.
Abstract:Despite incredible progress, many neural architectures fail to properly generalize beyond their training distribution. As such, learning to reason in a correct and generalizable way is one of the current fundamental challenges in machine learning. In this respect, logic puzzles provide a great testbed, as we can fully understand and control the learning environment. Thus, they allow to evaluate performance on previously unseen, larger and more difficult puzzles that follow the same underlying rules. Since traditional approaches often struggle to represent such scalable logical structures, we propose to model these puzzles using a graph-based approach. Then, we investigate the key factors enabling the proposed models to learn generalizable solutions in a reinforcement learning setting. Our study focuses on the impact of the inductive bias of the architecture, different reward systems and the role of recurrent modeling in enabling sequential reasoning. Through extensive experiments, we demonstrate how these elements contribute to successful extrapolation on increasingly complex puzzles.These insights and frameworks offer a systematic way to design learning-based systems capable of generalizable reasoning beyond interpolation.
Abstract:We are given a set of elements in a metric space. The distribution of the elements is arbitrary, possibly adversarial. Can we weigh the elements in a way that is resistant to such (adversarial) manipulations? This problem arises in various contexts. For instance, the elements could represent data points, requiring robust domain adaptation. Alternatively, they might represent tasks to be aggregated into a benchmark; or questions about personal political opinions in voting advice applications. This article introduces a theoretical framework for dealing with such problems. We propose clone-proof representation functions as a solution concept. These functions distribute importance across elements of a set such that similar objects (``clones'') share (some of) their weights, thus avoiding a potential bias introduced by their multiplicity. Our framework extends the maximum uncertainty principle to accommodate general metric spaces and includes a set of axioms - symmetry, continuity, and clone-proofness - that guide the construction of representation functions. Finally, we address the existence of representation functions satisfying our axioms in the significant case of Euclidean spaces and propose a general method for their construction.
Abstract:Information retrieval, specifically contract clause retrieval, is foundational to contract drafting because lawyers rarely draft contracts from scratch; instead, they locate and revise the most relevant precedent. We introduce the Atticus Clause Retrieval Dataset (ACORD), the first retrieval benchmark for contract drafting fully annotated by experts. ACORD focuses on complex contract clauses such as Limitation of Liability, Indemnification, Change of Control, and Most Favored Nation. It includes 114 queries and over 126,000 query-clause pairs, each ranked on a scale from 1 to 5 stars. The task is to find the most relevant precedent clauses to a query. The bi-encoder retriever paired with pointwise LLMs re-rankers shows promising results. However, substantial improvements are still needed to effectively manage the complex legal work typically undertaken by lawyers. As the first retrieval benchmark for contract drafting annotated by experts, ACORD can serve as a valuable IR benchmark for the NLP community.
Abstract:In the era of increasing privacy concerns and demand for personalized experiences, traditional Reinforcement Learning with Human Feedback (RLHF) frameworks face significant challenges due to their reliance on centralized data. We introduce Federated Reinforcement Learning with Human Feedback (FedRLHF), a novel framework that decentralizes the RLHF process. FedRLHF enables collaborative policy learning across multiple clients without necessitating the sharing of raw data or human feedback, thereby ensuring robust privacy preservation. Leveraging federated reinforcement learning, each client integrates human feedback locally into their reward functions and updates their policies through personalized RLHF processes. We establish rigorous theoretical foundations for FedRLHF, providing convergence guarantees, and deriving sample complexity bounds that scale efficiently with the number of clients. Empirical evaluations on the MovieLens and IMDb datasets demonstrate that FedRLHF not only preserves user privacy but also achieves performance on par with centralized RLHF, while enhancing personalization across diverse client environments.
Abstract:This work compares large language models (LLMs) and neuro-symbolic approaches in solving Raven's progressive matrices (RPM), a visual abstract reasoning test that involves the understanding of mathematical rules such as progression or arithmetic addition. Providing the visual attributes directly as textual prompts, which assumes an oracle visual perception module, allows us to measure the model's abstract reasoning capability in isolation. Despite providing such compositionally structured representations from the oracle visual perception and advanced prompting techniques, both GPT-4 and Llama-3 70B cannot achieve perfect accuracy on the center constellation of the I-RAVEN dataset. Our analysis reveals that the root cause lies in the LLM's weakness in understanding and executing arithmetic rules. As a potential remedy, we analyze the Abductive Rule Learner with Context-awareness (ARLC), a neuro-symbolic approach that learns to reason with vector-symbolic architectures (VSAs). Here, concepts are represented with distributed vectors s.t. dot products between encoded vectors define a similarity kernel, and simple element-wise operations on the vectors perform addition/subtraction on the encoded values. We find that ARLC achieves almost perfect accuracy on the center constellation of I-RAVEN, demonstrating a high fidelity in arithmetic rules. To stress the length generalization capabilities of the models, we extend the RPM tests to larger matrices (3x10 instead of typical 3x3) and larger dynamic ranges of the attribute values (from 10 up to 1000). We find that the LLM's accuracy of solving arithmetic rules drops to sub-10%, especially as the dynamic range expands, while ARLC can maintain a high accuracy due to emulating symbolic computations on top of properly distributed representations. Our code is available at https://github.com/IBM/raven-large-language-models.
Abstract:Several self-supervised learning (SSL) approaches have shown that redundancy reduction in the feature embedding space is an effective tool for representation learning. However, these methods consider a narrow notion of redundancy, focusing on pairwise correlations between features. To address this limitation, we formalize the notion of embedding space redundancy and introduce redundancy measures that capture more complex, higher-order dependencies. We mathematically analyze the relationships between these metrics, and empirically measure these redundancies in the embedding spaces of common SSL methods. Based on our findings, we propose Self Supervised Learning with Predictability Minimization (SSLPM) as a method for reducing redundancy in the embedding space. SSLPM combines an encoder network with a predictor engaging in a competitive game of reducing and exploiting dependencies respectively. We demonstrate that SSLPM is competitive with state-of-the-art methods and find that the best performing SSL methods exhibit low embedding space redundancy, suggesting that even methods without explicit redundancy reduction mechanisms perform redundancy reduction implicitly.