Probabilistic decoding in Large Language Models (LLMs) often yields inconsistent outputs, particularly on complex or long-form questions. Self-Consistency (SC) mitigates this for short-form QA by majority voting over exact strings, whereas Universal Self-Consistency (USC) and Weighted Unigram Consistency Score (WUCS) extend to long-form responses but lose accuracy on short-form benchmarks. We introduce Latent Self-Consistency (LSC), which selects the most semantically consistent response using learnable token embeddings. LSC's lightweight forward processing of summary tokens only introduces negligible runtime overhead (at most 0.9%) on top of standard decoding of the base LLM, and requires no changes to the model architecture. Across 6 short-form and 5 long-form reasoning benchmarks (e.g., MATH, MMLU, TruthfulQA), LSC surpasses SC, USC, and WUCS on both short-form and long-form on average performance, while adding negligible computational overhead on vanilla inference. These results position LSC as a reliable consistency-selection method that works effectively across various answer formats. Additionally, LSC provides well-calibrated confidence estimates, maintaining low expected calibration error across both answer formats.

Existing QA benchmarks typically assume distinct documents with minimal overlap, yet realworld retrieval-augmented generation (RAG) systems operate on corpora such as financial reports, legal codes, and patents, where information is highly redundant and documents exhibit strong inter-document similarity.This mismatch undermines evaluation validity: retrievers can be unfairly undervalued even when they retrieve documents that provide sufficient evidence, because redundancy across documents is not accounted for in evaluation.On the other hand, retrievers that perform well on standard benchmarks often generalize poorly to real-world corpora with highly similar and redundant documents.We present RARE (Redundancy-Aware Retrieval Evaluation), a framework for constructing realistic benchmarks by (i) decomposing documents into atomic facts to enable precise redundancy tracking and (ii) enhancing LLM-based data generation with CRRF.RAG benchmark data usually requires multiple quality criteria, but LLMs often yield trivial outputs.CRRF scores criteria separately and fuses decisions by rank, improving the reliability of generated data.Applying RARE to Finance, Legal, and Patent corpora, we introduce RedQA, where a strong retriever baseline drops from 66.4% PerfRecall@10 on 4-hop General-Wiki to 5.0-27.9%PerfRecall@10 at 4-hop depth, revealing robustness gaps that current benchmarks fail to capture.RARE enables practitioners to build domain-specific RAG evaluations that faithfully reflect real-world deployment conditions.

Examining logical inconsistencies among multiple statements (such as collections of sentences or question-answer pairs) is a crucial challenge in machine learning, particularly for ensuring the safety and reliability of models.Traditional methods that rely on 1:1 pairwise comparisons often fail to capture inconsistencies that only emerge when more than two statements are evaluated collectively.To address this gap, we introduce the task of setconsistency verification, an extension of natural language inference (NLI) that assesses the logical coherence of entire sets rather than isolated pairs.Building on this task, we present the Set-Consistency Energy Network (SC-Energy), a novel model that employs a margin-based loss to learn the compatibility among a collection of statements.Our approach not only efficiently verifies inconsistencies and pinpoints the specific statements responsible for logical contradictions, but also significantly outperforms existing methods, including prompting-based LLM models.Furthermore, we release two new datasets: Set-LConVQA and Set-SNLI for setconsistency verification task. Original Premise & Hypothesisp: A couple walk hand in hand down a street.h: A couple is walking together.

Reinforcement learning has shown great promise in aligning language models with human preferences in a variety of text generation tasks, including machine translation. For translation tasks, rewards can easily be obtained from quality estimation (QE) models which can generate rewards for unlabeled data. Despite its usefulness, reinforcement learning cannot exploit the gradients with respect to the QE score. We propose QE-EBM, a method of employing quality estimators as trainable loss networks that can directly backpropagate to the NMT model. We examine our method on several low and high resource target languages with English as the source language. QE-EBM outperforms strong baselines such as REINFORCE and proximal policy optimization (PPO) as well as supervised fine-tuning for all target languages, especially low-resource target languages. Most notably, for English-to-Mongolian translation, our method achieves improvements of 2.5 BLEU, 7.1 COMET-KIWI, 5.3 COMET, and 6.4 XCOMET relative to the supervised baseline.

Long-term time series forecasting poses significant challenges due to the complex dynamics and temporal variations, particularly when dealing with unseen patterns and data scarcity. Traditional supervised learning approaches, which rely on cleaned and labeled data, struggle to capture these unseen characteristics, limiting their effectiveness in real-world applications. In this study, we propose a semi-supervised approach that leverages multi-view setting on augmented data without requiring explicit future values as labels to address these limitations. By introducing a consensus promotion framework, our method enhances agreement among multiple single-view models on unseen augmented data. This approach not only improves forecasting accuracy but also mitigates error accumulation in long-horizon predictions. Furthermore, we explore the impact of autoregressive and non-autoregressive decoding schemes on error propagation, demonstrating the robustness of our model in extending prediction horizons. Experimental results show that our proposed method not only surpasses traditional supervised models in accuracy but also exhibits greater robustness when extending the prediction horizon.

Jiachen Zhao, Wenlong Zhao, Andrew Drozdov, Benjamin Rozonoyer, Md Arafat Sultan, Jay-Yoon Lee, Mohit Iyyer, Andrew McCallum. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The Retrieval Augmented Generation (RAG) framework utilizes a combination of parametric knowledge and external knowledge to demonstrate state-of-the-art performance on opendomain question answering tasks.However, the RAG framework suffers from performance degradation when the query is accompanied by irrelevant contexts.In this work, we propose the RE-RAG framework, which introduces a relevance estimator (RE) that not only provides relative relevance between contexts as previous rerankers did, but also provides confidence, which can be used to classify whether given context is useful for answering the given question.We propose a weakly supervised method for training the RE simply utilizing questionanswer data without any labels for correct contexts.We show that RE trained with a small generator (sLM) can not only improve the sLM fine-tuned together with RE but also improve previously unreferenced large language models (LLMs).Furthermore, we investigate new decoding strategies that utilize the proposed confidence measured by RE such as choosing to let the user know that it is "unanswerable" to answer the question given the retrieved contexts or choosing to rely on LLM's parametric knowledge rather than unrelated contexts.

A common retrieve-and-rerank paradigm involves retrieving relevant candidates from a broad set using a fast bi-encoder (BE), followed by applying expensive but accurate crossencoders (CE) to a limited candidate set.However, relying on this small subset is often susceptible to error propagation from the biencoders, which limits the overall performance.To address these issues, we propose the Comparing Multiple Candidates (CMC) framework.CMC compares a query and multiple embeddings of similar candidates (i.e., neighbors) through shallow self-attention layers, delivering rich representations contextualized to each other.Furthermore, CMC is scalable enough to handle multiple comparisons simultaneously.For example, comparing 10K candidates with CMC takes a similar amount of time as comparing 16 candidates with CE.Experimental results on the ZeSHEL dataset demonstrate that CMC, when plugged in between bi-encoders and cross-encoders as a seamless intermediate reranker (BE-CMC-CE), can effectively improve recall@k (+6.7%-p, +3.5%-p for R@16, R@64) compared to using only bi-encoders (BE-CE), with negligible slowdown (<7%).Additionally, to verify CMC's effectiveness as the final-stage reranker in improving top-1 accuracy, we conduct experiments on downstream tasks such as entity, passage, and dialogue ranking.The results indicate that CMC is not only faster (11x) but also often more effective than cross-encoders with improved prediction accuracy in Wikipedia entity linking (+0.7%-p) and DSTC7 dialogue ranking (+3.3%-p).

Abstract Retrieval Augmented Language Models (RALMs) have gained significant attention for their ability to generate accurate answers and improve efficiency. However, RALMs are inherently vulnerable to imperfect information due to their reliance on the imperfect retriever or knowledge source. We identify three common scenarios—unanswerable, adversarial, conflicting—where retrieved document sets can confuse RALMs with plausible real-world examples. We present the first comprehensive investigation to assess how well RALMs detect and handle such problematic scenarios. Among these scenarios, to systematically examine adversarial robustness we propose a new adversarial attack method, Generative model-based ADVersarial attack (GenADV) and a novel metric Robustness under Additional Document (RAD). Our findings reveal that RALMs often fail to identify the unanswerability or contradiction of a document set, which frequently leads to hallucinations. Moreover, we show that the addition of an adversary significantly degrades RALM’s performance, with the model becoming even more vulnerable when the two scenarios overlap (adversarial+ unanswerable). Our research identifies critical areas for assessing and enhancing the robustness of RALMs, laying the foundation for the development of more robust models.1

Belanger & McCallum (2016) and Gygli et al. (2017) have shown that an energy network can capture arbitrary dependencies amongst the output variables in structured prediction; however, their reliance on gradient-based inference (GBI) makes the inference slow and unstable. In this work, we propose Structured Energy As Loss (SEAL) to take advantage of the expressivity of energy networks without incurring the high inference cost. This is a novel learning framework that uses an energy network as a trainable loss function (loss-net) to train a separate neural network (task-net), which is then used to perform the inference through a forward pass. We establish SEAL as a general framework wherein various learning strategies like margin-based, regression, and noise-contrastive, could be employed to learn the parameters of loss-net. Through extensive evaluation on multi-label classification, semantic role labeling, and image segmentation, we demonstrate that SEAL provides various useful design choices, is faster at inference than GBI, and leads to significant performance gains over the baselines.

Vidhisha Balachandran, Artidoro Pagnoni, Jay Yoon Lee, Dheeraj Rajagopal, Jaime Carbonell, Yulia Tsvetkov. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

Rajarshi Das, Manzil Zaheer, Dung Thai, Ameya Godbole, Ethan Perez, Jay Yoon Lee, Lizhen Tan, Lazaros Polymenakos, Andrew McCallum. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Zhiyang Xu, Andrew Drozdov, Jay Yoon Lee, Tim O’Gorman, Subendhu Rongali, Dylan Finkbeiner, Shilpa Suresh, Mohit Iyyer, Andrew McCallum. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Multi-view learning makes use of diverse models arising from multiple sources of input or different feature subsets for the same task. For example, a given natural language processing task can combine evidence from models arising from character, morpheme, lexical, or phrasal views. The most common strategy with multi-view learning, especially popular in the neural network community, is to unify multiple representations into one unified vector through concatenation, averaging, or pooling, and then build a single-view model on top of the unified representation. As an alternative, we examine whether building one model per view and then unifying the different models can lead to improvements, especially in low-resource scenarios. More specifically, taking inspiration from co-training methods, we propose a semi-supervised learning approach based on multi-view models through consensus promotion, and investigate whether this improves overall performance. To test the multi-view hypothesis, we use moderately low-resource scenarios for nine languages and test the performance of the joint model for part-of-speech tagging and dependency parsing. The proposed model shows significant improvements across the test cases, with average gains of -0.9 ∼ +9.3 labeled attachment score (LAS) points. We also investigate the effect of unlabeled data on the proposed model by varying the amount of training data and by using different domains of unlabeled data.

Practitioners apply neural networks to increasingly complex problems in natural language processing, such as syntactic parsing and semantic role labeling that have rich output structures. Many such structured-prediction problems require deterministic constraints on the output values; for example, in sequence-to-sequence syntactic parsing, we require that the sequential outputs encode valid trees. While hidden units might capture such properties, the network is not always able to learn such constraints from the training data alone, and practitioners must then resort to post-processing. In this paper, we present an inference method for neural networks that enforces deterministic constraints on outputs without performing rule-based post-processing or expensive discrete search. Instead, in the spirit of gradient-based training, we enforce constraints with gradient-based inference (GBI): for each input at test-time, we nudge continuous model weights until the network’s unconstrained inference procedure generates an output that satisfies the constraints. We study the efficacy of GBI on three tasks with hard constraints: semantic role labeling, syntactic parsing, and sequence transduction. In each case, the algorithm not only satisfies constraints, but improves accuracy, even when the underlying network is stateof-the-art.

Neural models have shown several state-ofthe-art performances on Semantic Role Labeling (SRL).However, the neural models require an immense amount of semantic-role corpora and are thus not well suited for lowresource languages or domains.The paper proposes a semi-supervised semantic role labeling method that outperforms the state-ofthe-art in limited SRL training corpora.The method is based on explicitly enforcing syntactic constraints by augmenting the training objective with a syntactic-inconsistency loss component and uses SRL-unlabeled instances to train a joint-objective LSTM.On CoNLL-2012 English section, the proposed semi-supervised training with 1%, 10% SRLlabeled data and varying amounts of SRLunlabeled data achieves +1.58, +0.78 F1, respectively, over the pre-trained models that were trained on SOTA architecture with ELMo on the same SRL-labeled data.Additionally, by using the syntactic-inconsistency loss on inference time, the proposed model achieves +3.67, +2.1 F1 over pre-trained model on 1%, 10% SRL-labeled data, respectively.

This paper reports on methods and results of an applied research project by a team consisting of SAIC and four universities to develop, integrate, and evaluate new approaches to detect the weak signals characteristic of insider threats on organizations' information systems. Our system combines structural and semantic information from a real corporate database of monitored activity on their users' computers to detect independently developed red team inserts of malicious insider activities. We have developed and applied multiple algorithms for anomaly detection based on suspected scenarios of malicious insider behavior, indicators of unusual activities, high-dimensional statistical patterns, temporal sequences, and normal graph evolution. Algorithms and representations for dynamic graph processing provide the ability to scale as needed for enterprise-level deployments on real-time data streams. We have also developed a visual language for specifying combinations of features, baselines, peer groups, time periods, and algorithms to detect anomalies suggestive of instances of insider threat behavior. We defined over 100 data features in seven categories based on approximately 5.5 million actions per day from approximately 5,500 users. We have achieved area under the ROC curve values of up to 0.979 and lift values of 65 on the top 50 user-days identified on two months of real data.