arXiv subjects

Oren Kalinsky

Publications and source records attributed to Oren Kalinsky.

8 recordsLinked to original sources

arXiv2026

Knowing When to Quit: A Principled Framework for Dynamic Abstention in LLM Reasoning

LLMs utilizing chain-of-thought reasoning often waste substantial compute by producing long, incorrect responses. Abstention can mitigate this by withholding outputs unlikely to be correct. While most abstention methods decide to withhold outputs before or after generation, dynamic mid-generation abstention considers early termination of unpromising reasoning traces at each token position. Prior work has explored empirical variants of this idea, but principled guidance for the abstention rule remains lacking. We present a formal analysis of dynamic abstention for LLMs, modeling abstention as an explicit action within a regularized reinforcement learning framework. An abstention reward parameter controls the trade-off between compute and information. We show that abstaining when the value function falls below this reward strictly outperforms natural baselines under general conditions. We further derive a principled and efficient method to approximate the value function. Empirical results on mathematical reasoning and toxicity avoidance tasks support our theory and demonstrate improved selective accuracy over existing methods.

cs.LG↗

arXiv2026

HotelQuEST: Balancing Quality and Efficiency in Agentic Search

Agentic search has emerged as a promising paradigm for adaptive retrieval systems powered by large language models (LLMs). However, existing benchmarks primarily focus on quality, overlooking efficiency factors that are critical for real-world deployment. Moreover, real-world user queries often contain underspecified preferences, a challenge that remains largely underexplored in current agentic search evaluation. As a result, many agentic search systems remain impractical despite their impressive performance. In this work, we introduce HotelQuEST, a benchmark comprising 214 hotel search queries that range from simple factual requests to complex queries, enabling evaluation across the full spectrum of query difficulty. We further address the challenge of evaluating underspecified user preferences by collecting clarifications that make annotators' implicit preferences explicit for evaluation. We find that LLM-based agents achieve higher accuracy than traditional retrievers, but at substantially higher costs due to redundant tool calls and suboptimal routing that fails to match query complexity to model capability. Our analysis exposes inefficiencies in current agentic search systems and demonstrates substantial potential for cost-aware optimization.

cs.IR↗

arXiv2024

ChaI-TeA: A Benchmark for Evaluating Autocompletion of Interactions with LLM-based Chatbots

The rise of LLMs has deflected a growing portion of human-computer interactions towards LLM-based chatbots. The remarkable abilities of these models allow users to interact using long, diverse natural language text covering a wide range of topics and styles. Phrasing these messages is a time and effort consuming task, calling for an autocomplete solution to assist users. We introduce the task of chatbot interaction autocomplete. We present ChaI-TeA: CHat InTEraction Autocomplete; An autcomplete evaluation framework for LLM-based chatbot interactions. The framework includes a formal definition of the task, coupled with suitable datasets and metrics. We use the framework to evaluate After formally defining the task along with suitable datasets and metrics, we test 9 models on the defined auto completion task, finding that while current off-the-shelf models perform fairly, there is still much room for improvement, mainly in ranking of the generated suggestions. We provide insights for practitioners working on this task and open new research directions for researchers in the field. We release our framework to serve as a foundation for future research.

cs.CL↗

arXiv2024

Evaluating D-MERIT of Partial-annotation on Information Retrieval

Retrieval models are often evaluated on partially-annotated datasets. Each query is mapped to a few relevant texts and the remaining corpus is assumed to be irrelevant. As a result, models that successfully retrieve false negatives are punished in evaluation. Unfortunately, completely annotating all texts for every query is not resource efficient. In this work, we show that using partially-annotated datasets in evaluation can paint a distorted picture. We curate D-MERIT, a passage retrieval evaluation set from Wikipedia, aspiring to contain all relevant passages for each query. Queries describe a group (e.g., "journals about linguistics") and relevant passages are evidence that entities belong to the group (e.g., a passage indicating that "Language" is a journal about linguistics). We show that evaluating on a dataset containing annotations for only a subset of the relevant passages might result in misleading ranking of the retrieval systems and that as more relevant texts are included in the evaluation set, the rankings converge. We propose our dataset as a resource for evaluation and our study as a recommendation for balance between resource-efficiency and reliable evaluation when annotating evaluation sets for text retrieval.

cs.IR↗

arXiv2019

The TrieJax Architecture: Accelerating Graph Operations Through Relational Joins

Graph pattern matching (e.g., finding all cycles and cliques) has become an important component in many critical domains such as social networks, biology, and cyber-security. This development motivated research to develop faster algorithms that target graph pattern matching. In recent years, the database community has shown that mapping graph pattern matching problems to a new class of relational join algorithms provides an efficient framework for computing these problems. In this paper, we argue that this new class of relational join algorithms is highly amenable to specialized hardware acceleration thanks to two fundamental properties: improved locality and inherent concurrency. The improved locality is a result of the provably bound number of intermediate results these algorithms generate, which results in smaller working sets. In addition, their inherent concurrency can be leveraged for effective hardware acceleration and hiding memory latency. We demonstrate the hardware amenability of this new class of algorithms by introducing TrieJax, a hardware accelerator for graph pattern matching. The TrieJax design leverages the improved locality and high concurrency properties to dramatically accelerate graph pattern matching, and can be tightly integrated into existing manycore processors. We evaluate TrieJax on a set standard graph pattern matching queries and datasets. Our evaluation shows that TrieJax outperforms recently proposed hardware accelerators for graph and database processing that do not employ the new class of algorithms by 7-63x on average (up to 539x), while consuming 15-179x less energy (up to 1750x). systems that do incorporate modern relational join algorithms by 9-20x on average (up to 45x), while consuming 59-110x less energy (up to 372x).

cs.DB↗

arXiv2018

Efficiently Charting RDF

We propose a visual query language for interactively exploring large-scale knowledge graphs. Starting from an overview, the user explores bar charts through three interactions: class expansion, property expansion, and subject/object expansion. A major challenge faced is performance: a state-of-the-art SPARQL engine may require tens of minutes to compute the multiway join, grouping and counting required to render a bar chart. A promising alternative is to apply approximation through online aggregation, trading precision for performance. However, state-of-the-art online aggregation algorithms such as Wander Join have two limitations for our exploration scenario: (1) a high number of rejected paths slows the convergence of the count estimations, and (2) no unbiased estimator exists for counts under the distinct operator. We thus devise a specialized algorithm for online aggregation that augments Wander Join with exact partial computations to reduce the number of rejected paths encountered, as well as a novel estimator that we prove to be unbiased in the case of the distinct operator. In an experimental study with random interactions exploring two large-scale knowledge graphs, our algorithm shows a clear reduction in error with respect to computation time versus Wander Join.

cs.DB↗

arXiv2017

eLinda: Explorer for Linked Data

To realize the premise of the Semantic Web towards knowledgeable machines, one might often integrate an application with emerging RDF graphs. Nevertheless, capturing the content of a rich and open RDF graph by existing tools requires both time and expertise. We demonstrate eLinda - an explorer for Linked Data. The challenge addressed by eLinda is that of understanding the rich content of a given RDF graph. The core functionality is an exploration path, where each step produces a bar chart (histogram) that visualizes the distribution of classes in a set of nodes (URIs). In turn, each bar represents a set of nodes that can be further expanded through the bar chart in the path. We allow three types of explorations: subclass distribution, property distribution, and object distribution for a property of choice.

cs.DB↗

arXiv2016

Flexible Caching in Trie Joins

Traditional algorithms for multiway join computation are based on rewriting the order of joins and combining results of intermediate subqueries. Recently, several approaches have been proposed for algorithms that are "worst-case optimal" wherein all relations are scanned simultaneously. An example is Veldhuizen's Leapfrog Trie Join (LFTJ). An important advantage of LFTJ is its small memory footprint, due to the fact that intermediate results are full tuples that can be dumped immediately. However, since the algorithm does not store intermediate results, recurring joins must be reconstructed from the source relations, resulting in excessive memory traffic. In this paper, we address this problem by incorporating caches into LFTJ. We do so by adopting recent developments on join optimization, tying variable ordering to tree decomposition. While the traditional usage of tree decomposition computes the result for each bag in advance, our proposed approach incorporates caching directly into LFTJ and can dynamically adjust the size of the cache. Consequently, our solution balances memory usage and repeated computation, as confirmed by our experiments over SNAP datasets.

cs.DB↗