Alexander Bondarenko

Alexander Bondarenko, Denis Volk, Dmitrii Volkov et al.

We demonstrate LLM agent specification gaming by instructing models to win against a chess engine. We find reasoning models like OpenAI o3 and DeepSeek R1 will often hack the benchmark by default, while language models like GPT-4o and Claude 3.5 Sonnet need to be told that normal play won't work to hack. We improve upon prior work like (Hubinger et al., 2024; Meinke et al., 2024; Weij et al., 2024) by using realistic task prompts and avoiding excess nudging. Our results suggest reasoning models may resort to hacking to solve difficult problems, as observed in OpenAI (2024)'s o1 Docker escape during cyber capabilities testing.

Alexander Bondarenko, Adrian Viehweger

The retrieval-augmented generation (RAG) approach is used to reduce the confabulation of large language models (LLMs) for question answering by retrieving and providing additional context coming from external knowledge sources (e.g., by adding the context to the prompt). However, injecting incorrect information can mislead the LLM to generate an incorrect answer. In this paper, we evaluate the effectiveness and robustness of four LLMs against misinformation - Gemma 2, GPT-4o-mini, Llama~3.1, and Mixtral - in answering biomedical questions. We assess the answer accuracy on yes-no and free-form questions in three scenarios: vanilla LLM answers (no context is provided), "perfect" augmented generation (correct context is provided), and prompt-injection attacks (incorrect context is provided). Our results show that Llama 3.1 (70B parameters) achieves the highest accuracy in both vanilla (0.651) and "perfect" RAG (0.802) scenarios. However, the accuracy gap between the models almost disappears with "perfect" RAG, suggesting its potential to mitigate the LLM's size-related effectiveness differences. We further evaluate the ability of the LLMs to generate malicious context on one hand and the LLM's robustness against prompt-injection attacks on the other hand, using metrics such as attack success rate (ASR), accuracy under attack, and accuracy drop. As adversaries, we use the same four LLMs (Gemma 2, GPT-4o-mini, Llama 3.1, and Mixtral) to generate incorrect context that is injected in the target model's prompt. Interestingly, Llama is shown to be the most effective adversary, causing accuracy drops of up to 0.48 for vanilla answers and 0.63 for "perfect" RAG across target models. Our analysis reveals that robustness rankings vary depending on the evaluation measure, highlighting the complexity of assessing LLM resilience to adversarial attacks.

Michael Völske, Alexander Bondarenko, Maik Fröbe et al.

Recently, neural networks have been successfully employed to improve upon state-of-the-art performance in ad-hoc retrieval tasks via machine-learned ranking functions. While neural retrieval models grow in complexity and impact, little is understood about their correspondence with well-studied IR principles. Recent work on interpretability in machine learning has provided tools and techniques to understand neural models in general, yet there has been little progress towards explaining ranking models. We investigate whether one can explain the behavior of neural ranking models in terms of their congruence with well understood principles of document ranking by using established theories from axiomatic IR. Axiomatic analysis of information retrieval models has formalized a set of constraints on ranking decisions that reasonable retrieval models should fulfill. We operationalize this axiomatic thinking to reproduce rankings based on combinations of elementary constraints. This allows us to investigate to what extent the ranking decisions of neural rankers can be explained in terms of retrieval axioms, and which axioms apply in which situations. Our experimental study considers a comprehensive set of axioms over several representative neural rankers. While the existing axioms can already explain the particularly confident ranking decisions rather well, future work should extend the axiom set to also cover the other still "unexplainable" neural IR rank decisions.

Matthias Schildwächter, Alexander Bondarenko, Julian Zenker et al.

We present CAM (comparative argumentative machine), a novel open-domain IR system to argumentatively compare objects with respect to information extracted from the Common Crawl. In a user study, the participants obtained 15% more accurate answers using CAM compared to a "traditional" keyword-based search and were 20% faster in finding the answer to comparative questions.

Alexander Panchenko, Alexander Bondarenko, Mirco Franzek et al.

We tackle the tasks of automatically identifying comparative sentences and categorizing the intended preference (e.g., "Python has better NLP libraries than MATLAB" => (Python, better, MATLAB). To this end, we manually annotate 7,199 sentences for 217 distinct target item pairs from several domains (27% of the sentences contain an oriented comparison in the sense of "better" or "worse"). A gradient boosting model based on pre-trained sentence embeddings reaches an F1 score of 85% in our experimental evaluation. The model can be used to extract comparative sentences for pro/con argumentation in comparative / argument search engines or debating technologies.