General-purpose open-domain dense retrieval systems are usually trained with a large, eclectic mix of corpora and search tasks. How should these diverse corpora and tasks be sampled for training? Conventional approaches sample them uniformly, proportional to their instance population sizes, or depend on human-level expert supervision. It is well known that the training data sampling strategy can greatly impact model performance. However, how to find the optimal strategy has not been adequately studied in the context of embedding models. We propose **Inf-DDS**, a novel reinforcement learning–driven sampling framework that adaptively reweighs training datasets guided by influence-based reward signals and is much more lightweight w.r.t. GPU consumption. Our technique iteratively refines the sampling policy, prioritizing datasets that maximize model performance on a target development set. We evaluate the efficacy of our sampling strategy on a wide range of text retrieval tasks, demonstrating strong improvements in retrieval performance and better adaptation compared to existing gradient-based sampling methods, while also being 1.5—4× cheaper in GPU compute. Our sampling strategy achieves a 5.03 absolute NDCG@10 improvement while training the multilingual bge-m3 model and an absolute NDCG@10 improvement of 0.94 while training all-MiniLM-L6-v2, even when starting from expert-assigned weights on a large pool of training datasets.

Code explanations are increasingly generated by large language models and used in software engineering workflows, making reliable evaluation essential. However, existing model-based and embedding-based methods often fail to distinguish correct explanations from partially or fully incorrect ones, and their similarity scores are poorly calibrated and do not reflect meaningful differences in explanation quality. To address this, we propose ODASim(Orderly, Dstinctive, and Absolute Similarity), a model-agnostic graded fine-tuning framework for embedding models that learns calibrated similarity representations between code and explanations. To support fine-grained supervision and evaluation, we also introduce ODA-X, a novel benchmark for code-to-explanation quality grading, comprising code–explanation pairs graded similarity labels derived from strategic perturbations of gold explanations. We apply our ODASim approach to multiple embedding models and evaluate it on two benchmarks: widely popular CodeXGLUE and our proposed benchmark ODA-X, spanning four programming languages - Python, Java, JavaScript, and Go. Results show that our method achieves up to 35% improvement in F1 score and 85% reduction in Expected Calibration Error (ECE), enabling reliable evaluation of code to explanation quality.

As language models are increasingly deployed as autonomous agents, understanding whether they can anticipate the consequences of their actions is critical. We introduce RUBICON, a benchmark that tests whether models recognize when their own actions have made a task impossible to complete. Three procedurally generated puzzle domains measure this ability across spatial, resource management, and temporal reasoning, with a symbolic oracle providing ground truth. Evaluating eleven models, we find that all but one almost never flag a doomed situation on their own (0–7% detection vs. 83% for GPT-5.4), even though they correctly identify the same situations up to 100% of the time when asked directly. Models whose reasoning traces contain correct unsolvability analysis frequently override it to provide a move anyway: self-override rates reach 34–77% for models with high detection ability. When asked to identify which past decision caused the failure, models fail to trace back to the responsible action, instead blaming the moment the problem became obvious rather than the earlier decision that caused it. These findings highlight a fundamental gap: the ability to detect danger does not translate into the ability to avoid it.

LLMs acquire vast amounts of knowledge during pre-training, but often lack the specialized knowledge needed to answer questions from niche sources such as manuals or technical documents unseen during pre-training. Continued pre-training (CPT) is widely used to inject such knowledge into model parameters. However, niche documents seldom repeat facts, making it difficult for CPT to robustly acquire such knowledge. Recent works address this by generating multiple paraphrases of the new knowledge, but paraphrasing is computationally expensive and typically requires powerful LLMs. In this work, we introduce KItCAT: Knowledge Injection via Corrupted Auto-regressive Training, a lightweight training strategy that reduces the need for paraphrasing in decoder-only LLMs. KItCAT augments standard next-token prediction by stochastically corrupting the input sequence. During training, a random subset of input tokens is replaced with other vocabulary tokens while the original next-token labels are kept unchanged. This simple intervention generates diverse training inputs from each sample, enabling large-scale data augmentation at negligible cost. We show that KItCAT consistently improves over CPT across multiple datasets and model families. Code is available at https://github.com/meghanadhpulivarthi/KItCAT.

Evaluating Large Language Models (LLMs) in low-resource and linguistically diverse languages remains a significant challenge in NLP, particularly for languages using non-Latin scripts like those spoken in India. Existing benchmarks predominantly focus on English, leaving substantial gaps in assessing LLM capabilities in these languages. We introduce MILU, a Multi-task Indic Language Understanding Benchmark, a comprehensive evaluation benchmark designed to address this gap. MILU spans 8 domains and 41 subjects across 11 Indic languages, reflecting general and culturally specific knowledge. With an India-centric design, incorporates material from regional and state-level examinations, covering topics such as local history, arts, festivals, and laws, alongside standard subjects like science and mathematics. We evaluate over 42 LLMs, and find that current LLMs struggle with MILU, with GPT-4o achieving the highest average accuracy at 74 percent. Open multilingual models outperform language-specific fine-tuned models, which perform only slightly better than random baselines. Models also perform better in high resource languages as compared to low resource ones. Domain-wise analysis indicates that models perform poorly in culturally relevant areas like Arts and Humanities, Law and Governance compared to general fields like STEM. To the best of our knowledge, MILU is the first of its kind benchmark focused on Indic languages, serving as a crucial step towards comprehensive cultural evaluation. All code, benchmarks, and artifacts are publicly available to foster open research.

Given the large number of Hindi speakers worldwide, there is a pressing need for robust and efficient information retrieval systems for Hindi. Despite ongoing research, comprehensive benchmarks for evaluating retrieval models in Hindi are lacking. To address this gap, we introduce the Hindi-BEIR benchmark, comprising 15 datasets across seven distinct tasks. We evaluate state-of-the-art multilingual retrieval models on the Hindi-BEIR benchmark, identifying task and domain-specific challenges that impact Hindi retrieval performance. Building on the insights from these results, we introduce NLLB-E5, a multilingual retrieval model that leverages a zero-shot approach to support Hindi without the need for Hindi training data. We believe our contributions, including the release of the Hindi-BEIR benchmark and the NLLB-E5 model, will be a valuable resource for researchers and promote advancements in multilingual retrieval models.

Large Language Models (LLMs) perform well on unseen tasks in English, but their abilities in non-English languages are less explored due to limited benchmarks and training data. To bridge this gap, we introduce the Indic-QA Benchmark, a large dataset for context-grounded question answering in 11 major Indian languages, covering both extractive and abstractive tasks. Evaluations of multilingual LLMs, including instruction fine-tuned versions, revealed weak performance in low-resource languages due to a strong English-language bias in their training data. We also investigated the Translate-Test paradigm,where inputs are translated to English for processing and the results are translated back into the source language for output. This approach outperformed multilingual LLMs, particularly in low-resource settings. By releasing Indic-QA, we aim to promote further research into LLMs’ question-answering capabilities in low-resource languages. This benchmark offers a critical resource to address existing limitations and foster multilingual understanding.

The two-stage retrieval paradigm has gained popularity, where a neural model serves as a re-ranker atop a non-neural first-stage retriever. We argue that this approach, involving two disparate models without interaction, represents a suboptimal choice. To address this, we propose a unified encoder-decoder architecture with a novel training regimen which enables the encoder representation to be used for retrieval and the decoder for re-ranking within a single unified model, facilitating end-to-end retrieval. We incorporate XTR-style retrieval on top of the trained MonoT5 reranker to specifically concentrate on addressing practical constraints to create a lightweight model. Results on the BIER benchmark demonstrate the effectiveness of our unified architecture, featuring a highly optimized index and parameters. It outperforms ColBERT, XTR, and even serves as a superior re-ranker compared to the Mono-T5 reranker. The performance gains of our proposed system in reranking become increasingly evident as model capacity grows, particularly when compared to rerankers operating over traditional first-stage retrievers like BM25. This is encouraging, as it suggests that we can integrate more advanced retrievers to further enhance final reranking performance. In contrast, BM25’s static nature limits its potential for such improvements.

XTR (Lee et al., 2023) introduced an efficient multi-vector retrieval method that addresses the limitations of the ColBERT (Khattab and Zaharia, 2020model by simplifying retrieval into a single stage through a modified learning objective. While XTR eliminates the need for multistage retrieval, it doesn’t incorporate the efficiency optimizations from ColBERTv2 (Santhanam et al., 2022, which improve indexing and retrieval speed. In this work, we enhance XTR by integrating ColBERTv2’s optimizations, showing that the combined approach preserves the strengths of both models. This results in a more efficient and scalable solution for multi-vector retrieval, while maintaining XTR’s streamlined retrieval process.

Question answering (QA) over tables and linked text, also called TextTableQA, has witnessed significant research in recent years, as tables are often found embedded in documents along with related text. HybridQA and OTT-QA are the two best-known TextTableQA datasets, with questions that are best answered by combining information from both table cells and linked text passages. A common challenge in both datasets, and TextTableQA in general, is that the training instances include just the question and answer, where the gold answer may match not only multiple table cells across table rows but also multiple text spans within the scope of a table row and its associated text. This leads to a noisy multi-instance training regime. We present MITQA, a transformer-based TextTableQA system that is explicitly designed to cope with distant supervision along both these axes, through a multi-instance loss objective, together with careful curriculum design. Our experiments show that the proposed multi-instance distant supervision approach helps MITQA get sate-of-the-art results beating the existing baselines for both HybridQA and OTT-QA, putting MITQA at the top of HybridQA leaderboard with best EM and F1 scores on a held out test set.

The field of Question Answering (QA) has made remarkable progress in recent years, thanks to the advent of large pre-trained language models, newer realistic benchmark datasets with leaderboards, and novel algorithms for key components such as retrievers and readers. In this paper, we introduce PrimeQA: a one-stop and open-source QA repository with an aim to democratize QA research and facilitate easy replication of state-of-the-art (SOTA) QA methods. PrimeQA supports core QA functionalities like retrieval and reading comprehension as well as auxiliary capabilities such as question generation. It has been designed as an end-to-end toolkit for various use cases: building front-end applications, replicating SOTA methods on public benchmarks, and expanding pre-existing methods. PrimeQA is available at: https://github.com/primeqa.

Table Question Answering (Table QA) systems have been shown to be highly accurate when trained and tested on open-domain datasets built on top of Wikipedia tables. However, it is not clear whether their performance remains the same when applied to domain-specific scientific and business documents, encountered in industrial settings, which exhibit some unique characteristics: (a) they contain tables with a much more complex layout than Wikipedia tables (including hierarchical row and column headers), (b) they contain domain-specific terms, and (c) they are typically not accompanied by domain-specific labeled data that can be used to train Table QA models. To understand the performance of Table QA approaches in this setting, we introduce AIT-QA; a domain-specific Table QA test dataset. While focusing on the airline industry, AIT-QA reflects the challenges that domain-specific documents pose to Table QA, outlined above. In this work, we describe the creation of the dataset and report zero-shot experimental results of three SOTA Table QA methods. The results clearly expose the limitations of current methods with a best accuracy of just 51.8%. We also present pragmatic table pre-processing steps to pivot and project complex tables into a layout suitable for the SOTA Table QA models. Finally, we provide data-driven insights on how different aspects of this setting (including hierarchical headers, domain-specific terminology, and paraphrasing) affect Table QA methods, in order to help the community develop improved methods for domain-specific Table QA.

Weakly-supervised table question-answering (TableQA) models have achieved state-of-art performance by using pre-trained BERT transformer to jointly encoding a question and a table to produce structured query for the question. However, in practical settings TableQA systems are deployed over table corpora having topic and word distributions quite distinct from BERT’s pretraining corpus. In this work we simulate the practical topic shift scenario by designing novel challenge benchmarks WikiSQL-TS and WikiTable-TS, consisting of train-dev-test splits in five distinct topic groups, based on the popular WikiSQL and WikiTable-Questions datasets. We empirically show that, despite pre-training on large open-domain text, performance of models degrades significantly when they are evaluated on unseen topics. In response, we propose T3QA (Topic Transferable Table Question Answering) a pragmatic adaptation framework for TableQA comprising of: (1) topic-specific vocabulary injection into BERT, (2) a novel text-to-text transformer generator (such as T5, GPT2) based natural language question generation pipeline focused on generating topic-specific training data, and (3) a logical form re-ranker. We show that T3QA provides a reasonably good baseline for our topic shift benchmarks. We believe our topic split benchmarks will lead to robust TableQA solutions that are better suited for practical deployment

Natural Language Query interfaces allow the end-users to access the desired information without the need to know any specialized query language, data storage, or schema details. Even with the recent advances in NLP research space, the state-of-the-art QA systems fall short of understanding implicit intents of real-world Business Intelligence (BI) queries in enterprise systems, since Natural Language Understanding still remains an AI-hard problem. We posit that deploying ontology reasoning over domain semantics can help in achieving better natural language understanding for QA systems. In this paper, we specifically focus on building a Schema Aware Semantic Reasoning Framework that translates natural language interpretation as a sequence of solvable tasks by an ontology reasoner. We apply our framework on top of an ontology based, state-of-the-art natural language question-answering system ATHENA, and experiment with 4 benchmarks focused on BI queries. Our experimental numbers empirically show that the Schema Aware Semantic Reasoning indeed helps in achieving significantly better results for handling BI queries with an average accuracy improvement of ~30%