Daniel M. Bikel

Sander Land, Daniel M. Bikel

LLM benchmark labels are frozen at release and silently propagated into downstream benchmarks, errors and all. We introduce an Item Response Theory-based indicator that surfaces likely mislabels at 95% precision in the top 200 examples across seven preference and multiple-choice benchmarks using responses from 114 models, outperforming a supervised classifier. We trace these errors to mechanical labeling heuristics, upstream annotation mistakes inherited unchanged from source datasets, and fundamentally ambiguous items without a defensible single label. The same model fit reveals that reward models specialize in stylistic preference rather than factual knowledge, and identifies one frontier reward model that agrees with detected mislabels at 78% accuracy versus 38% for its peers, consistent with benchmark contamination or benchmark-specific over-optimization.

Hsuan Su, Rebecca Qian, Chinnadhurai Sankar et al. · meta-ai, mila

Recent works have shown considerable improvements in task-oriented dialogue (TOD) systems by utilizing pretrained large language models (LLMs) in an end-to-end manner. However, the biased behavior of each component in a TOD system and the error propagation issue in the end-to-end framework can lead to seriously biased TOD responses. Existing works of fairness only focus on the total bias of a system. In this paper, we propose a diagnosis method to attribute bias to each component of a TOD system. With the proposed attribution method, we can gain a deeper understanding of the sources of bias. Additionally, researchers can mitigate biased model behavior at a more granular level. We conduct experiments to attribute the TOD system's bias toward three demographic axes: gender, age, and race. Experimental results show that the bias of a TOD system usually comes from the response generation model.

Zhe Dong, Jianmo Ni, Daniel M. Bikel et al.

Dual encoders have been used for question-answering (QA) and information retrieval (IR) tasks with good results. Previous research focuses on two major types of dual encoders, Siamese Dual Encoder (SDE), with parameters shared across two encoders, and Asymmetric Dual Encoder (ADE), with two distinctly parameterized encoders. In this work, we explore different ways in which the dual encoder can be structured, and evaluate how these differences can affect their efficacy in terms of QA retrieval tasks. By evaluating on MS MARCO, open domain NQ and the MultiReQA benchmarks, we show that SDE performs significantly better than ADE. We further propose three different improved versions of ADEs by sharing or freezing parts of the architectures between two encoder towers. We find that sharing parameters in projection layers would enable ADEs to perform competitively with or outperform SDEs. We further explore and explain why parameter sharing in projection layer significantly improves the efficacy of the dual encoders, by directly probing the embedding spaces of the two encoder towers with t-SNE algorithm.

Yiming Zhang, Jianfeng Chi, Hailey Nguyen et al.

Text generation has a fundamental limitation almost by definition: there is no taking back tokens that have been generated, even when they are clearly problematic. In the context of language model safety, when a partial unsafe generation is produced, language models by their nature tend to happily keep on generating similarly unsafe additional text. This is in fact how safety alignment of frontier models gets circumvented in the wild, despite great efforts in improving their safety. Deviating from the paradigm of approaching safety alignment as prevention (decreasing the probability of harmful responses), we propose backtracking, a technique that allows language models to "undo" and recover from their own unsafe generation through the introduction of a special [RESET] token. Our method can be incorporated into either SFT or DPO training to optimize helpfulness and harmlessness. We show that models trained to backtrack are consistently safer than baseline models: backtracking Llama-3-8B is four times more safe than the baseline model (6.1\% $\to$ 1.5\%) in our evaluations without regression in helpfulness. Our method additionally provides protection against four adversarial attacks including an adaptive attack, despite not being trained to do so.

Yi-Lin Tuan, Xilun Chen, Eric Michael Smith et al.

As large language models (LLMs) become easily accessible nowadays, the trade-off between safety and helpfulness can significantly impact user experience. A model that prioritizes safety will cause users to feel less engaged and assisted while prioritizing helpfulness will potentially cause harm. Possible harms include teaching people how to build a bomb, exposing youth to inappropriate content, and hurting users' mental health. In this work, we propose to balance safety and helpfulness in diverse use cases by controlling both attributes in LLM. We explore training-free and fine-tuning methods that do not require extra human annotations and analyze the challenges of controlling safety and helpfulness in LLMs. Our experiments demonstrate that our method can rewind a learned model and unlock its controllability.

Zhenyu Zhao, Aparna Balagopalan, Adi Agrawal et al.

Given the increased use of LLMs in financial systems today, it becomes important to evaluate the safety and robustness of such systems. One failure mode that LLMs frequently display in general domain settings is that of sycophancy. That is, models prioritize agreement with expressed user beliefs over correctness, leading to decreased accuracy and trust. In this work, we focus on evaluating sycophancy that LLMs display in agentic financial tasks. Our findings are three-fold: first, we find the models show only low to modest drops in performance in the face of user rebuttals or contradictions to the reference answer, which distinguishes sycophancy that models display in financial agentic settings from findings in prior work. Second, we introduce a suite of tasks to test for sycophancy by user preference information that contradicts the reference answer and find that most models fail in the presence of such inputs. Lastly, we benchmark different modes of recovery such as input filtering with a pretrained LLM.

Matthew Finlayson, Ilia Kulikov, Daniel M. Bikel et al. · meta-ai

Large language models (LLMs) often struggle with knowledge intensive NLP tasks, such as answering "Who won the latest World Cup?" because the knowledge they learn during training may be insufficient or outdated. Conditioning generation on retrieved documents -- a technique known as retrieval augmented generation (RAG) -- mitigates these shortcomings by allowing the model to leverage in-context information. Practitioners can improve LLM RAG performance by fine-tuning on retrieval-augmented instructions, but must beware that this can cause undesirable model behaviors like hallucinations. We attribute this degradation to the fact that the training data is likely to be out-of-distribution for the model and may suffer from quality issues, such as misalignment between retrievals and target responses (since retrievals are frequently added post-hoc). We propose a recipe for training RAG-enabled LLMs using self-generated demonstrations, thereby avoiding training on out-of-distribution text and integrating retrievals into the LLM responses. We evaluate our method on knowledge intensive question answering (QA) tasks and show that our method teaches LLMs to properly handle in-context retrievals and abstain from questions it will likely get wrong. Compared to conventional RA-IT methods, our method prevents model degradation in non-RAG settings while exhibiting superior QA performance.

Nicholas FitzGerald, Jan A. Botha, Daniel Gillick et al.

We present an instance-based nearest neighbor approach to entity linking. In contrast to most prior entity retrieval systems which represent each entity with a single vector, we build a contextualized mention-encoder that learns to place similar mentions of the same entity closer in vector space than mentions of different entities. This approach allows all mentions of an entity to serve as "class prototypes" as inference involves retrieving from the full set of labeled entity mentions in the training set and applying the nearest mention neighbor's entity label. Our model is trained on a large multilingual corpus of mention pairs derived from Wikipedia hyperlinks, and performs nearest neighbor inference on an index of 700 million mentions. It is simpler to train, gives more interpretable predictions, and outperforms all other systems on two multilingual entity linking benchmarks.

Oshin Agarwal, Daniel M. Bikel

Entity Linking has two main open areas of research: 1) generate candidate entities without using alias tables and 2) generate more contextual representations for both mentions and entities. Recently, a solution has been proposed for the former as a dual-encoder entity retrieval system (Gillick et al., 2019) that learns mention and entity representations in the same space, and performs linking by selecting the nearest entity to the mention in this space. In this work, we use this retrieval system solely for generating candidate entities. We then rerank the entities by using a cross-attention encoder over the target mention and each of the candidate entities. Whereas a dual encoder approach forces all information to be contained in the small, fixed set of vector dimensions used to represent mentions and entities, a crossattention model allows for the use of detailed information (read: features) from the entirety of each <mention, context, candidate entity> tuple. We experiment with features used in the reranker including different ways of incorporating document-level context. We achieve state-of-the-art results on TACKBP-2010 dataset, with 92.05% accuracy. Furthermore, we show how the rescoring model generalizes well when trained on the larger CoNLL-2003 dataset and evaluated on TACKBP-2010.