Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models

This paper explores the impact of extending input lengths on the capabilities of Large Language Models (LLMs). Despite LLMs advancements in recent times, their performance consistency across different input lengths is not well understood. We investigate this aspect by introducing a novel QA reasoning framework, specifically designed to assess the impact of input length. We isolate the effect of input length using multiple versions of the same sample, each being extended with padding of different lengths, types and locations. Our findings show a notable degradation in LLMs' reasoning performance at much shorter input lengths than their technical maximum. We show that the degradation trend appears in every version of our dataset, although at different intensities. Additionally, our study reveals that traditional perplexity metrics do not correlate with performance of LLMs' in long input reasoning tasks. We analyse our results and identify failure modes that can serve as useful guides for future research, potentially informing strategies to address the limitations observed in LLMs.

Transpose Attack: Stealing Datasets with Bidirectional Training

Deep neural networks are normally executed in the forward direction. However, in this work, we identify a vulnerability that enables models to be trained in both directions and on different tasks. Adversaries can exploit this capability to hide rogue models within seemingly legitimate models. In addition, in this work we show that neural networks can be taught to systematically memorize and retrieve specific samples from datasets. Together, these findings expose a novel method in which adversaries can exfiltrate datasets from protected learning environments under the guise of legitimate models. We focus on the data exfiltration attack and show that modern architectures can be used to secretly exfiltrate tens of thousands of samples with high fidelity, high enough to compromise data privacy and even train new models. Moreover, to mitigate this threat we propose a novel approach for detecting infected models.

Guiding LLM to Fool Itself: Automatically Manipulating Machine Reading Comprehension Shortcut Triggers

Recent applications of LLMs in Machine Reading Comprehension (MRC) systems have shown impressive results, but the use of shortcuts, mechanisms triggered by features spuriously correlated to the true label, has emerged as a potential threat to their reliability. We analyze the problem from two angles: LLMs as editors, guided to edit text to mislead LLMs; and LLMs as readers, who answer questions based on the edited text. We introduce a framework that guides an editor to add potential shortcuts-triggers to samples. Using GPT4 as the editor, we find it can successfully edit trigger shortcut in samples that fool LLMs. Analysing LLMs as readers, we observe that even capable LLMs can be deceived using shortcut knowledge. Strikingly, we discover that GPT4 can be deceived by its own edits (15% drop in F1). Our findings highlight inherent vulnerabilities of LLMs to shortcut manipulations. We publish ShortcutQA, a curated dataset generated by our framework for future research.

Clever Hans or Neural Theory of Mind? Stress Testing Social Reasoning in Large Language Models

The escalating debate on AI's capabilities warrants developing reliable metrics to assess machine "intelligence". Recently, many anecdotal examples were used to suggest that newer large language models (LLMs) like ChatGPT and GPT-4 exhibit Neural Theory-of-Mind (N-ToM); however, prior work reached conflicting conclusions regarding those abilities. We investigate the extent of LLMs' N-ToM through an extensive evaluation on 6 tasks and find that while LLMs exhibit certain N-ToM abilities, this behavior is far from being robust. We further examine the factors impacting performance on N-ToM tasks and discover that LLMs struggle with adversarial examples, indicating reliance on shallow heuristics rather than robust ToM abilities. We caution against drawing conclusions from anecdotal examples, limited benchmark testing, and using human-designed psychological tests to evaluate models.

Transferability Ranking of Adversarial Examples

Adversarial examples can be used to maliciously and covertly change a model's prediction. It is known that an adversarial example designed for one model can transfer to other models as well. This poses a major threat because it means that attackers can target systems in a blackbox manner. In the domain of transferability, researchers have proposed ways to make attacks more transferable and to make models more robust to transferred examples. However, to the best of our knowledge, there are no works which propose a means for ranking the transferability of an adversarial example in the perspective of a blackbox attacker. This is an important task because an attacker is likely to use only a select set of examples, and therefore will want to select the samples which are most likely to transfer. In this paper we suggest a method for ranking the transferability of adversarial examples without access to the victim's model. To accomplish this, we define and estimate the expected transferability of a sample given limited information about the victim. We also explore practical scenarios: where the adversary can select the best sample to attack and where the adversary must use a specific sample but can choose different perturbations. Through our experiments, we found that our ranking method can increase an attacker's success rate by up to 80% compared to the baseline (random selection without ranking).