Oren Kurland

Sagie Dekel, Moshe Tennenholtz, Oren Kurland

Retrieval Augmented Generation (RAG) is a highly effective paradigm for keeping LLM-based responses up-to-date and reducing the likelihood of hallucinations. Yet, RAG was recently shown to be quite vulnerable to corpus knowledge poisoning: an attacker injects misleading documents to the corpus to steer an LLM's output to an undesired response. We argue that the standard causal attention mechanism in LLMs enables harmful cross-document interactions, specifically in cases of attacks. Accordingly, we introduce a novel defense approach for RAG: Sparse Document Attention RAG (SDAG). This is a block-sparse attention mechanism that disallows cross-attention between retrieved documents. SDAG requires a minimal inference-time change to the attention mask; furthermore, no fine-tuning or additional architectural changes are needed. We present an empirical evaluation of LLM-based question answering (QA) with a variety of attack strategies on RAG. We show that our SDAG method substantially outperforms the standard causal attention mechanism in terms of attack success rate. We further demonstrate the clear merits of integrating SDAG with state-of-the-art RAG defense methods. Specifically, the integration results in performance that is statistically significantly better than the state-of-the-art.

Tommy Mordo, Sagie Dekel, Omer Madmon et al.

Competitive search is a setting where document publishers modify them to improve their ranking in response to a query. Recently, publishers have increasingly leveraged LLMs to generate and modify competitive content. We introduce Reinforcement Learning from Ranker Feedback (RLRF), a framework that trains LLMs using preference datasets derived from ranking competitions. The goal of a publisher (LLM-based) agent is to optimize content for improved ranking while accounting for the strategies of competing agents. We generate the datasets using approaches that do not rely on human-authored data. We show that our proposed agents consistently and substantially outperform previously suggested approaches for LLM-based competitive document modification. We further show that our agents are effective with ranking functions they were not trained for (i.e., out of distribution) and they adapt to strategic opponents. These findings provide support to the significant potential of using reinforcement learning in competitive search.

Gregory Goren, Oren Kurland, Moshe Tennenholtz et al.

In competitive search settings such as the Web, many documents' authors (publishers) opt to have their documents highly ranked for some queries. To this end, they modify the documents - specifically, their content - in response to induced rankings. Thus, the search engine affects the content in the corpus via its ranking decisions. We present a first study of the ability of search engines to drive pre-defined, targeted, content effects in the corpus using simple techniques. The first is based on the herding phenomenon - a celebrated result from the economics literature - and the second is based on biasing the relevance ranking function. The types of content effects we study are either topical or touch on specific document properties - length and inclusion of query terms. Analysis of ranking competitions we organized between incentivized publishers shows that the types of content effects we target can indeed be attained by applying our suggested techniques. These findings have important implications with regard to the role of search engines in shaping the corpus.

Ziv Vasilisky, Moshe Tennenholtz, Oren Kurland

The ranking incentives of many authors of Web pages play an important role in the Web dynamics. That is, authors who opt to have their pages highly ranked for queries of interest, often respond to rankings for these queries by manipulating their pages; the goal is to improve the pages' future rankings. Various theoretical aspects of this dynamics have recently been studied using game theory. However, empirical analysis of the dynamics is highly constrained due to lack of publicly available datasets.We present an initial such dataset that is based on TREC's ClueWeb09 dataset. Specifically, we used the WayBack Machine of the Internet Archive to build a document collection that contains past snapshots of ClueWeb documents which are highly ranked by some initial search performed for ClueWeb queries. Temporal analysis of document changes in this dataset reveals that findings recently presented for small-scale controlled ranking competitions between documents' authors also hold for Web data. Specifically, documents' authors tend to mimic the content of documents that were highly ranked in the past, and this practice can result in improved ranking.

Gregory Goren, Oren Kurland, Moshe Tennenholtz et al.

The Web is a canonical example of a competitive retrieval setting where many documents' authors consistently modify their documents to promote them in rankings. We present an automatic method for quality-preserving modification of document content -- i.e., maintaining content quality -- so that the document is ranked higher for a query by a non-disclosed ranking function whose rankings can be observed. The method replaces a passage in the document with some other passage. To select the two passages, we use a learning-to-rank approach with a bi-objective optimization criterion: rank promotion and content-quality maintenance. We used the approach as a bot in content-based ranking competitions. Analysis of the competitions demonstrates the merits of our approach with respect to human content modifications in terms of rank promotion, content-quality maintenance and relevance.

Eilon Sheetrit, Anna Shtok, Oren Kurland

According to common relevance-judgments regimes, such as TREC's, a document can be deemed relevant to a query even if it contains a very short passage of text with pertinent information. This fact has motivated work on passage-based document retrieval: document ranking methods that induce information from the document's passages. However, the main source of passage-based information utilized was passage-query similarities. We address the challenge of utilizing richer sources of passage-based information to improve document retrieval effectiveness. Specifically, we devise a suite of learning-to-rank-based document retrieval methods that utilize an effective ranking of passages produced in response to the query; the passage ranking is also induced using a learning-to-rank approach. Some of the methods quantify the ranking of the passages of a document. Others utilize the feature-based representation of passages used for learning a passage ranker. Empirical evaluation attests to the clear merits of our methods with respect to highly effective baselines. Our best performing method is based on learning a document ranking function using document-query features and passage-query features of the document's passage most highly ranked.

Gregory Goren, Oren Kurland, Moshe Tennenholtz et al.

For many queries in the Web retrieval setting there is an on-going ranking competition: authors manipulate their documents so as to promote them in rankings. Such competitions can have unwarranted effects not only in terms of retrieval effectiveness, but also in terms of ranking robustness. A case in point, rankings can (rapidly) change due to small indiscernible perturbations of documents. While there has been a recent growing interest in analyzing the robustness of classifiers to adversarial manipulations, there has not yet been a study of the robustness of relevance-ranking functions. We address this challenge by formally analyzing different definitions and aspects of the robustness of learning-to-rank-based ranking functions. For example, we formally show that increased regularization of linear ranking functions increases ranking robustness. This finding leads us to conjecture that decreased variance of any ranking function results in increased robustness. We propose several measures for quantifying ranking robustness and use them to analyze ranking competitions between documents' authors. The empirical findings support our formal analysis and conjecture for both RankSVM and LambdaMART.

Oren Kurland, Eyal Krikon

Exploiting information induced from (query-specific) clustering of top-retrieved documents has long been proposed as a means for improving precision at the very top ranks of the returned results. We present a novel language model approach to ranking query-specific clusters by the presumed percentage of relevant documents that they contain. While most previous cluster ranking approaches focus on the cluster as a whole, our model utilizes also information induced from documents associated with the cluster. Our model substantially outperforms previous approaches for identifying clusters containing a high relevant-document percentage. Furthermore, using the model to produce document ranking yields precision-at-top-ranks performance that is consistently better than that of the initial ranking upon which clustering is performed. The performance also favorably compares with that of a state-of-the-art pseudo-feedback-based retrieval method.

Anna Khudyak Kozorovitsky, Oren Kurland

Methods for fusing document lists that were retrieved in response to a query often utilize the retrieval scores and/or ranks of documents in the lists. We present a novel fusion approach that is based on using, in addition, information induced from inter-document similarities. Specifically, our methods let similar documents from different lists provide relevance-status support to each other. We use a graph-based method to model relevance-status propagation between documents. The propagation is governed by inter-document-similarities and by retrieval scores of documents in the lists. Empirical evaluation demonstrates the effectiveness of our methods in fusing TREC runs. The performance of our most effective methods transcends that of effective fusion methods that utilize only retrieval scores or ranks.