Roger Wattenhofer’s research while affiliated with ETH Zurich and other places

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.

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.

We introduce the Byzantine Selection Problem, living at the intersection of game theory and fault-tolerant distributed computing. Here, an event organizer is presented with a group of n agents, and wants to select $\ell < n$ of them to form a team. For these purposes, each agent i self-reports a positive skill value $v_i$, and a team's value is the sum of its members' skill values. Ideally, the value of the team should be as large as possible, which can be easily achieved by selecting agents with the highest $\ell$ skill values. However, an unknown subset of at most $t < n$ agents are byzantine and hence not to be trusted, rendering their true skill values as 0. In the spirit of the distributed computing literature, the identity of the byzantine agents is not random but instead chosen by an adversary aiming to minimize the value of the chosen team. Can we still select a team with good guarantees in this adversarial setting? As it turns out, deterministically, it remains optimal to select agents with the highest $\ell$ values. Yet, if $t \geq \ell$, the adversary can choose to make all selected agents byzantine, leading to a team of value zero. To provide meaningful guarantees, one hence needs to allow for randomization, in which case the expected value of the selected team needs to be maximized, assuming again that the adversary plays to minimize it. For this case, we provide linear-time randomized algorithms that maximize the expected value of the selected team.

We consider voting on multiple independent binary issues. In addition, a weighting vector for each voter defines how important they consider each issue. The most natural way to aggregate the votes into a single unified proposal is issue-wise majority (IWM): taking a majority opinion for each issue. However, in a scenario known as Ostrogorski's Paradox, an IWM proposal may not be a Condorcet winner, or it may even fail to garner majority support in a special case known as Anscombe's Paradox. We show that it is co-NP-hard to determine whether there exists a Condorcet-winning proposal even without weights. In contrast, we prove that the single-switch condition provides an Ostrogorski-free voting domain under identical weighting vectors. We show that verifying the condition can be achieved in linear time and no-instances admit short, efficiently computable proofs in the form of forbidden substructures. On the way, we give the simplest linear-time test for the voter/candidate-extremal-interval condition in approval voting and the simplest and most efficient algorithm for recognizing single-crossing preferences in ordinal voting. We then tackle Anscombe's Paradox. Under identical weight vectors, we can guarantee a majority-supported proposal agreeing with IWM on strictly more than half of the overall weight, while with two distinct weight vectors, such proposals can get arbitrarily far from IWM. The severity of such examples is controlled by the maximum average topic weight $\tilde{w}_{max}$: a simple bound derived from a partition-based approach is tight on a large portion of the range $\tilde{w}_{max} \in (0,1)$. Finally, we extend Wagner's rule to the weighted setting: an average majority across topics of at least $\frac{3}{4}$'s precludes Anscombe's paradox from occurring.

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.

Consider a graph $G = (V, E)$ and some commuters, each specified by a tuple $(u, v, b)$ consisting of two nodes in the graph $u, v \in V$ and a non-negative real number b, specifying their budget. The goal is to find a pricing function p of the edges of G that maximizes the revenue generated by the commuters. Here, each commuter $(u, v, b)$ either pays the lowest-cost of a u-v path under the pricing p, or 0, if this exceeds their budget b. We study this problem for the case where G is a bounded-width grid graph and give a polynomial-time approximation algorithm with approximation ratio $O(\log |E|)$. Our approach combines existing ideas with new insights. Most notably, we employ a rather seldom-encountered technique that we coin under the name 'assume-implement dynamic programming.' This technique involves dynamic programming where some information about the future decisions of the dynamic program is guessed in advance and 'assumed' to hold, and then subsequent decisions are forced to 'implement' the guess. This enables computing the cost of the current transition by using information that would normally only be available in the future.

Given the low throughput of blockchains like Bitcoin and Ethereum, scalability -- the ability to process an increasing number of transactions -- has become a central focus of blockchain research. One promising approach is the parallelization of transaction execution across multiple threads. However, achieving efficient parallelization requires a redesign of the incentive structure within the fee market. Currently, the fee market does not differentiate between transactions that access multiple high-demand resources versus a single low-demand one, as long as they require the same computational effort. Addressing this discrepancy is crucial for enabling more effective parallel execution. In this work, we aim to bridge the gap between the current fee market and the need for parallel execution by exploring alternative fee market designs. To this end, we propose a framework consisting of two key components: a Gas Computation Mechanism (GCM), which quantifies the load a transaction places on the network in terms of parallelization and computation, measured in units of gas, and a Transaction Fee Mechanism (TFM), which assigns a price to each unit of gas. We also introduce a set of desirable properties for a GCM, present multiple candidate mechanisms, and evaluate them against the properties. One promising candidate emerges: the weighted area GCM. Notably, this mechanism can be seamlessly composed with existing TFMs, such as EIP-1559. While our exploration primarily focuses on the execution component of the fee, which directly relates to parallel execution, we also outline how it could be integrated with fees associated with other factors, such as storage and data bandwidth, by drawing a parallel to a multi-dimensional fee market.

In stable matching, one must find a matching between two sets of agents, commonly men and women, or job applicants and job positions. Each agent has a preference ordering over who they want to be matched with. Moreover a matching is said to be stable if no pair of matched agents prefer each other compared to their current matching. We consider solving stable matching in a distributed synchronous setting, where each agent is its own process. Moreover, we assume up to $t_L$ agents on one side and $t_R$ on the other side can be byzantine. After properly defining the stable matching problem in this setting, we study its solvability. When there are as many agents on each side with fully-ordered preference lists, we give necessary and sufficient conditions for stable matching to be solvable in the synchronous setting. These conditions depend on the communication model used, i.e., if parties on the same side are allowed to communicate directly, and on the presence of a cryptographic setup, i.e., digital signatures.

In this work, we investigate the computational aspects of asynchronous cellular automata (ACAs), a modification of cellular automata in which cells update independently, following an asynchronous schedule. We introduce flip automata networks (FAN), a simple modification of automata networks that remain robust under any asynchronous update schedule. We show that asynchronous automata can efficiently simulate their synchronous counterparts with a linear memory overhead, which improves upon the previously established quadratic bound. Additionally, we address the universality gap for (a)synchronous cellular automata -- the boundary separating universal and non-universal automata, which is still not fully understood. We tighten this boundary by proving that all one-way asynchronous automata lack universal computational power. Conversely, we establish the existence of a universal 6-state first-neighbor automaton in one dimension and a 3-state von Neumann automaton in two dimensions, which represent the smallest known universal constructions to date.

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.