Graph foundation models (GFMs) have recently gained significant attention. However, the unique data processing and evaluation setups employed by different studies hinder a deeper understanding of their progress. Additionally, current research tends to focus on specific subsets of graph learning tasks, such as structural tasks, node-level tasks, or classification tasks. As a result, they often incorporate specialized modules tailored to particular task types, losing their applicability to other graph learning tasks and contradicting the original intent of foundation models to be universal. Therefore, to enhance consistency, coverage, and diversity across domains, tasks, and research interests within the graph learning community in the evaluation of GFMs, we propose GFMBench-a systematic and comprehensive benchmark comprising 26 datasets. Moreover, we introduce LangGFM, a novel GFM that relies entirely on large language models. By revisiting and exploring the effective graph textualization principles, as well as repurposing successful techniques from graph augmentation and graph self-supervised learning within the language space, LangGFM achieves performance on par with or exceeding the state of the art across GFMBench, which can offer us new perspectives, experiences, and baselines to drive forward the evolution of GFMs.

Recently, Large Language Models (LLMs) and Vision Language Models (VLMs) have demonstrated aptitude as potential substitutes for human participants in experiments testing psycholinguistic phenomena. However, an understudied question is to what extent models that only have access to vision and text modalities are able to implicitly understand sound-based phenomena via abstract reasoning from orthography and imagery alone. To investigate this, we analyse the ability of VLMs and LLMs to demonstrate sound symbolism (i.e., to recognise a non-arbitrary link between sounds and concepts) as well as their ability to "hear" via the interplay of the language and vision modules of open and closed-source multimodal models. We perform multiple experiments, including replicating the classic Kiki-Bouba and Mil-Mal shape and magnitude symbolism tasks and comparing human judgements of linguistic iconicity with that of LLMs. Our results show that VLMs demonstrate varying levels of agreement with human labels, Figure 1: Illustration of the 3 main experiments we and more task information may be required perform. Firstly, Shape Symbolism is a binary choice for VLMs versus their human counterparts for between two pseudowords to best describe an object that is in silico experimentation. We additionally see spiky or rounded. Magnitude Symbolism involves a binary through higher maximum agreement levels that choice between two pseudowords to best describe an object Magnitude Symbolism is an easier pattern for that is small or large. Finally, Iconicity involves rating VLMs to identify than Shape Symbolism, and the perceived iconicity of words, or how much their written/phonetic that an understanding of linguistic iconicity is form is representative of what they describe.

Syllogistic reasoning is crucial for Natural Language Inference (NLI). This capability is particularly significant in specialized domains such as biomedicine, where it can support automatic evidence interpretation and scientific discovery. This paper presents SylloBio-NLI, a novel framework that leverages external ontologies to systematically instantiate diverse syllogistic arguments for biomedical NLI. We employ SylloBio-NLI to evaluate Large Language Models (LLMs) on identifying valid conclusions and extracting supporting evidence across 28 syllogistic schemes instantiated with human genome pathways. Extensive experiments reveal that biomedical syllogistic reasoning is particularly challenging for zero-shot LLMs, which achieve an average accuracy between 70% on generalized modus ponens and 23% on disjunctive syllogism. At the same time, we found that few-shot prompting can boost the performance of different LLMs, including Gemma (+14%) and LLama-3 (+43%). However, a deeper analysis shows that both techniques exhibit high sensitivity to superficial lexical variations, highlighting a dependency between reliability, models' architecture, and pre-training regime. Overall, our results indicate that, while in-context examples have the potential to elicit syllogistic reasoning in LLMs, existing models are still far from achieving the robustness and consistency required for safe biomedical NLI applications.

Training-free embedding methods directly leverage pretrained large language models (LLMs) to embed text, bypassing the costly and complex procedure of contrastive learning. Previous training-free embedding methods have mainly focused on optimizing embedding prompts and have overlooked the benefits of utilizing the generative abilities of LLMs. We propose a novel method, GenEOL, which uses LLMs to generate diverse transformations of a sentence that preserve its meaning, and aggregates the resulting embeddings of these transformations to enhance the overall sentence embedding. GenEOL significantly outperforms the existing training-free embedding methods by an average of 2.85 points across several LLMs on the sentence semantic text similarity (STS) benchmark. Our analysis shows that GenEOL stabilizes representation quality across LLM layers and is robust to perturbations of embedding prompts. GenEOL also achieves notable gains on multiple clustering, reranking and pair-classification tasks from the MTEB benchmark.

We develop a method that integrates the tree of thoughts and multi-agent framework to enhance the capability of pre-trained language models in solving complex, unfamiliar games. The method decomposes game-solving into four incremental tasks -- game summarization, area selection, action extraction, and action validation -- each assigned to a specific language-model agent. By constructing a tree of thoughts, the method simulates reasoning paths and allows agents to collaboratively distill game representations and tactics, mitigating the limitations of language models in reasoning and long-term memorization. Additionally, an automated fine-tuning process further optimizes the agents' performance by ranking query-response pairs based on game outcomes, e.g., winning or losing. We apply the method to a non-cooperative game and demonstrate a 65 percent winning rate against benchmark algorithms, with an additional 10 percent improvement after fine-tuning. In contrast to existing deep learning algorithms for game solving that require millions of training samples, the proposed method consumes approximately 1000 training samples, highlighting its efficiency and scalability.

Motion time series collected from mobile and wearable devices such as smartphones and smartwatches offer significant insights into human behavioral patterns, with wide applications in healthcare, automation, IoT, and AR/XR due to their low-power, always-on nature. However, given security and privacy concerns, building large-scale motion time series datasets remains difficult, preventing the development of pre-trained models for human activity analysis. Typically, existing models are trained and tested on the same dataset, leading to poor generalizability across variations in device location, device mounting orientation and human activity type. In this paper, we introduce UniMTS, the first unified pre-training procedure for motion time series that generalizes across diverse device latent factors and activities. Specifically, we employ a contrastive learning framework that aligns motion time series with text descriptions enriched by large language models. This helps the model learn the semantics of time series to generalize across activities. Given the absence of large-scale motion time series data, we derive and synthesize time series from existing motion skeleton data with all-joint coverage. Spatio-temporal graph networks are utilized to capture the relationships across joints for generalization across different device locations. We further design rotation-invariant augmentation to make the model agnostic to changes in device mounting orientations. Our model shows exceptional generalizability across 18 motion time series classification benchmark datasets, outperforming the best baselines by 340% in the zero-shot setting, 16.3% in the few-shot setting, and 9.2% in the full-shot setting.

One strength of modern language models is their ability to incorporate information from a user-input context when answering queries. However, they are not equally sensitive to the subtle changes to that context. To quantify this, Du et al. (2024) gives an information-theoretic metric to measure such sensitivity. Their metric, susceptibility, is defined as the degree to which contexts can influence a model's response to a query at a distributional level. However, exactly computing susceptibility is difficult and, thus, Du et al. (2024) falls back on a Monte Carlo approximation. Due to the large number of samples required, the Monte Carlo approximation is inefficient in practice. As a faster alternative, we propose Fisher susceptibility, an efficient method to estimate the susceptibility based on Fisher information. Empirically, we validate that Fisher susceptibility is comparable to Monte Carlo estimated susceptibility across a diverse set of query domains despite its being $70\times$ faster. Exploiting the improved efficiency, we apply Fisher susceptibility to analyze factors affecting the susceptibility of language models. We observe that larger models are as susceptible as smaller ones.

Large Language Models (LLMs) have shown significant potential in designing reward functions for Reinforcement Learning (RL) tasks. However, obtaining high-quality reward code often involves human intervention, numerous LLM queries, or repetitive RL training. To address these issues, we propose CARD, a LLM-driven Reward Design framework that iteratively generates and improves reward function code. Specifically, CARD includes a Coder that generates and verifies the code, while a Evaluator provides dynamic feedback to guide the Coder in improving the code, eliminating the need for human feedback. In addition to process feedback and trajectory feedback, we introduce Trajectory Preference Evaluation (TPE), which evaluates the current reward function based on trajectory preferences. If the code fails the TPE, the Evaluator provides preference feedback, avoiding RL training at every iteration and making the reward function better aligned with the task objective. Empirical results on Meta-World and ManiSkill2 demonstrate that our method achieves an effective balance between task performance and token efficiency, outperforming or matching the baselines across all tasks. On 10 out of 12 tasks, CARD shows better or comparable performance to policies trained with expert-designed rewards, and our method even surpasses the oracle on 3 tasks.

Automated fact verification plays an essential role in fostering trust in the digital space. Despite the growing interest, the verification of temporal facts has not received much attention in the community. Temporal fact verification brings new challenges where cues of the temporal information need to be extracted and temporal reasoning involving various temporal aspects of the text must be applied. In this work, we propose an end-to-end solution for temporal fact verification that considers the temporal information in claims to obtain relevant evidence sentences and harness the power of large language model for temporal reasoning. Recognizing that temporal facts often involve events, we model these events in the claim and evidence sentences. We curate two temporal fact datasets to learn time-sensitive representations that encapsulate not only the semantic relationships among the events, but also their chronological proximity. This allows us to retrieve the top-k relevant evidence sentences and provide the context for a large language model to perform temporal reasoning and outputs whether a claim is supported or refuted by the retrieved evidence sentences. Experiment results demonstrate that the proposed approach significantly enhances the accuracy of temporal claim verification, thereby advancing current state-of-the-art in automated fact verification.

Sparse autoencoders (SAEs) are a promising technique for decomposing language model activations into interpretable linear features. However, current SAEs fall short of completely explaining model performance, resulting in "dark matter": unexplained variance in activations. This work investigates dark matter as an object of study in its own right. Surprisingly, we find that much of SAE dark matter--about half of the error vector itself and >90% of its norm--can be linearly predicted from the initial activation vector. Additionally, we find that the scaling behavior of SAE error norms at a per token level is remarkably predictable: larger SAEs mostly struggle to reconstruct the same contexts as smaller SAEs. We build on the linear representation hypothesis to propose models of activations that might lead to these observations, including postulating a new type of "introduced error"; these insights imply that the part of the SAE error vector that cannot be linearly predicted ("nonlinear" error) might be fundamentally different from the linearly predictable component. To validate this hypothesis, we empirically analyze nonlinear SAE error and show that 1) it contains fewer not yet learned features, 2) SAEs trained on it are quantitatively worse, 3) it helps predict SAE per-token scaling behavior, and 4) it is responsible for a proportional amount of the downstream increase in cross entropy loss when SAE activations are inserted into the model. Finally, we examine two methods to reduce nonlinear SAE error at a fixed sparsity: inference time gradient pursuit, which leads to a very slight decrease in nonlinear error, and linear transformations from earlier layer SAE outputs, which leads to a larger reduction.