Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP states, and Attention Consistency Training (AttCT), which matches per-head attention distributions. Second, we apply consistency training to four additional safety threats: persona in-context learning attacks, adversarial frustration, prefill attacks, and conditional misalignment. Across several models and threat settings, we find that consistency training reduces misalignment well beyond the sycophancy and jailbreak settings studied in prior work. The effective consistency target depends on the structure of the threat: representation-level targets suit wrapper-style failures, while output-level BCT is the reliable choice for trajectory-level failures such as prefill, frustration, and conditional misalignment. We also find cases of cross-threat generalization, where training against one failure mode improves robustness to another, and identify a shared residual-stream mechanism underlying ACT, MLPCT, and AttCT, while distinguishing BCT as mechanistically distinct. Our results suggest that consistency training is a flexible and extensible framework for alignment, capable of unifying defenses against a broader class of model pathologies.

Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model. Across five model families and seven bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out (LODO) transfer, and causal intervention. The susceptibility is largely shaped by alignment tuning rather than pretraining: pretrained base models generally cave much less to these biases, and their activations carry much weaker cue-specific signal beyond question content. Within aligned models, each bias has a coherent linear direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases do not collapse into a single shared representation, however: cross-bias overlap is model-specific, and even behaviorally similar biases occupy different directions. The same intervention also provides a proof-of-concept debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers across all instruct families. Cue-induced bias is therefore best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.

We study how explicit world-modeling objectives affect the internal representations and downstream capability of Transformers, using Rubik’s Cubes as our training domain. We ask: (1) how does explicitly pretraining a world model affect a model’s latent representations, (2) how does world-model quality affect post-training performance, and (3) how should a finite data budget be split between pretraining and task-specific fine-tuning? We compare standard action-prediction fine-tuning with two strategies that add explicit state-prediction supervision: state pretraining followed by fine-tuning, and joint action and state training. We measure task accuracy after Group Relative Policy Optimization (GRPO). We further find that explicit world-modeling yields better representations in terms of higher probing accuracy and steerability of the model, and that better representations yield larger gains from GRPO, especially on harder cube states. Finally, when the fine-tuning budget is held fixed, probe accuracy strongly predicts the GRPO improvement. Under a fixed total data budget, accuracy is maximized by allocating only a small fraction to pretraining.

In-Context Learning (ICL) is a technique by which language models make predictions based on examples provided in their input context. Previously, their context window size imposed a limit on the number of examples that can be shown, making example selection techniques crucial for identifying the maximally effective set of examples. However, the recent advent of Long Context Language Models (LCLMs) has significantly increased the number of examples that can be included in context, raising an important question of whether ICL performance in a many-shot regime is still sensitive to the method of sample selection. To answer this, we revisit these approaches in the context of LCLMs through extensive experiments on 18 datasets spanning 4 tasks. Surprisingly, we observe that sophisticated example selection techniques do not yield significant improvements over a simple random sample selection method. Instead, we discover that the advent of LCLMs has fundamentally shifted the challenge of ICL from that of selecting the most effective examples to that of collecting sufficient examples to fill the context window. Specifically, in certain datasets, including all available examples does not fully utilize the context window; however, by augmenting the examples in context with a simple data augmentation approach, we substantially improve ICL performance by 5%.

We present a generalizable classification approach that leverages Large Language Models (LLMs) to facilitate the detection of implicitly encoded social meaning in conversations. We design a multi-faceted prompt to extract a textual explanation of the reasoning that connects visible cues to underlying social meanings. These extracted explanations or rationales serve as augmentations to the conversational text to facilitate dialogue understanding and transfer. Our empirical results over 2,340 experimental settings demonstrate the significant positive impact of adding these rationales. Our findings hold true for in-domain classification, zero-shot, and few-shot domain transfer for two different social meaning detection tasks, each spanning two different corpora.

Dialogue models are able to generate coherent and fluent responses, but they can still be challenging to control and may produce non-engaging, unsafe results. This unpredictability diminishes user trust and can hinder the use of the models in the real world. To address this, we introduce DialGuide, a novel framework for controlling dialogue model behavior using natural language rules, or guidelines. These guidelines provide information about the context they are applicable to and what should be included in the response, allowing the models to generate responses that are more closely aligned with the developer’s expectations and intent. We evaluate DialGuide on three tasks in open-domain dialogue response generation: guideline selection, response generation, and response entailment verification. Our dataset contains 10,737 positive and 15,467 negative dialogue context-response-guideline triplets across two domains - chit-chat and safety. We provide baseline models for the tasks and benchmark their performance. We also demonstrate that DialGuide is effective in the dialogue safety domain, producing safe and engaging responses that follow developer guidelines.

While large neural-based conversational models have become increasingly proficient dialogue agents, recent work has highlighted safety issues with these systems. For example, these systems can be goaded into generating toxic content, often perpetuating social biases or stereotypes. We investigate a retrieval-based approach for reducing bias and toxicity in responses from chatbots. It uses in-context learning to steer a model towards safer generations. Concretely, to generate a response to an unsafe dialogue context, we retrieve demonstrations of safe responses to similar dialogue contexts. We find our method performs competitively with existing approaches to dialogue safety without requiring training. We also show, using automatic and human evaluation, that reductions in toxicity obtained using our approach are not at the cost engagingness or coherency. Finally, we note our method can be used in compliment to existing dialogue safety approaches, such as RLHF.

While the NLP community has produced numerous summarization benchmarks, none provide the rich annotations required to simultaneously address many important problems related to control and reliability. We introduce a Wikipedia-derived benchmark, complemented by a rich set of crowd-sourced annotations, that supports 8 interrelated tasks: (i) extractive summarization; (ii) abstractive summarization; (iii) topic-based summarization; (iv) compressing selected sentences into a one-line summary; (v) surfacing evidence for a summary sentence; (vi) predicting the factual accuracy of a summary sentence; (vii) identifying unsubstantiated spans in a summary sentence; (viii) correcting factual errors in summaries. We compare various methods on this benchmark and discover that on multiple tasks, moderately-sized fine-tuned models consistently outperform much larger few-shot prompted language models. For factuality-related tasks, we also evaluate existing heuristics to create training data and find that training on them results in worse performance than training on 20× less human-labeled data. Our articles draw from 6 domains, facilitating cross-domain analysis. On some tasks, the amount of training data matters more than the domain where it comes from, while for other tasks training specifically on data from the target domain, even if limited, is more beneficial.

Target-guided response generation enables dialogue systems to smoothly transition a conversation from a dialogue context toward a target sentence. Such control is useful for designing dialogue systems that direct a conversation toward specific goals, such as creating non-obtrusive recommendations or introducing new topics in the conversation. In this paper, we introduce a new technique for target-guided response generation, which first finds a bridging path of commonsense knowledge concepts between the source and the target, and then uses the identified bridging path to generate transition responses. Additionally, we propose techniques to re-purpose existing dialogue datasets for target-guided generation. Experiments reveal that the proposed techniques outperform various baselines on this task. Finally, we observe that the existing automated metrics for this task correlate poorly with human judgement ratings. We propose a novel evaluation metric that we demonstrate is more reliable for target-guided response evaluation. Our work generally enables dialogue system designers to exercise more control over the conversations that their systems produce.

Instruction tuning is an emergent paradigm in NLP wherein natural language instructions are leveraged with language models to induce zero-shot performance on unseen tasks. Dialogue is an especially interesting area in which to explore instruction tuning because dialogue systems perform multiple kinds of tasks related to language (e.g., natural language understanding and generation, domain-specific interaction), yet instruction tuning has not been systematically explored for dialogue-related tasks. We introduce InstructDial, an instruction tuning framework for dialogue, which consists of a repository of 48 diverse dialogue tasks in a unified text-to-text format created from 59 openly available dialogue datasets. We explore cross-task generalization ability on models tuned on InstructDial across diverse dialogue tasks. Our analysis reveals that InstructDial enables good zero-shot performance on unseen datasets and tasks such as dialogue evaluation and intent detection, and even better performance in a few-shot setting. To ensure that models adhere to instructions, we introduce novel meta-tasks. We establish benchmark zero-shot and few-shot performance of models trained using the proposed framework on multiple dialogue tasks.

Fact-checking is an essential tool to mitigate the spread of misinformation and disinformation. We introduce the task of fact-checking in dialogue, which is a relatively unexplored area. We construct DialFact, a testing benchmark dataset of 22,245 annotated conversational claims, paired with pieces of evidence from Wikipedia. There are three sub-tasks in DialFact: 1) Verifiable claim detection task distinguishes whether a response carries verifiable factual information; 2) Evidence retrieval task retrieves the most relevant Wikipedia snippets as evidence; 3) Claim verification task predicts a dialogue response to be supported, refuted, or not enough information. We found that existing fact-checking models trained on non-dialogue data like FEVER fail to perform well on our task, and thus, we propose a simple yet data-efficient solution to effectively improve fact-checking performance in dialogue. We point out unique challenges in DialFact such as handling the colloquialisms, coreferences, and retrieval ambiguities in the error analysis to shed light on future research in this direction.

Dialogue systems pretrained with large language models generate locally coherent responses, but lack fine-grained control over responses necessary to achieve specific goals. A promising method to control response generation is exemplar-based generation, in which models edit exemplar responses that are retrieved from training data, or hand-written to strategically address discourse-level goals, to fit new dialogue contexts. We present an Exemplar-based Dialogue Generation model, EDGE, that uses the semantic frames present in exemplar responses to guide response generation. We show that controlling dialogue generation based on the semantic frames of exemplars improves the coherence of generated responses, while preserving semantic meaning and conversation goals present in exemplar responses.

The advent of contextual word embeddings — representations of words which incorporate semantic and syntactic information from their context—has led to tremendous improvements on a wide variety of NLP tasks. However, recent contextual models have prohibitively high computational cost in many use-cases and are often hard to interpret. In this work, we demonstrate that our proposed distillation method, which is a simple extension of CBOW-based training, allows to significantly improve computational efficiency of NLP applications, while outperforming the quality of existing static embeddings trained from scratch as well as those distilled from previously proposed methods. As a side-effect, our approach also allows a fair comparison of both contextual and static embeddings via standard lexical evaluation tasks.

Large-scale models for learning fixed-dimensional cross-lingual sentence representations like LASER (Artetxe and Schwenk, 2019b) lead to significant improvement in performance on downstream tasks. However, further increases and modifications based on such large-scale models are usually impractical due to memory limitations. In this work, we introduce a lightweight dual-transformer architecture with just 2 layers for generating memory-efficient cross-lingual sentence representations. We explore different training tasks and observe that current cross-lingual training tasks leave a lot to be desired for this shallow architecture. To ameliorate this, we propose a novel cross-lingual language model, which combines the existing single-word masked language model with the newly proposed cross-lingual token-level reconstruction task. We further augment the training task by the introduction of two computationally-lite sentence-level contrastive learning tasks to enhance the alignment of cross-lingual sentence representation space, which compensates for the learning bottleneck of the lightweight transformer for generative tasks. Our comparisons with competing models on cross-lingual sentence retrieval and multilingual document classification confirm the effectiveness of the newly proposed training tasks for a shallow model.

The aim of this paper is to mitigate the shortcomings of automatic evaluation of open-domain dialog systems through multi-reference evaluation. Existing metrics have been shown to correlate poorly with human judgement, particularly in open-domain dialog. One alternative is to collect human annotations for evaluation, which can be expensive and time consuming. To demonstrate the effectiveness of multi-reference evaluation, we augment the test set of DailyDialog with multiple references. A series of experiments show that the use of multiple references results in improved correlation between several automatic metrics and human judgement for both the quality and the diversity of system output.

We study the problem of generating interesting endings for stories. Neural generative models have shown promising results for various text generation problems. Sequence to Sequence (Seq2Seq) models are typically trained to generate a single output sequence for a given input sequence. However, in the context of a story, multiple endings are possible. Seq2Seq models tend to ignore the context and generate generic and dull responses. Very few works have studied generating diverse and interesting story endings for the same story context. In this paper, we propose models which generate more diverse and interesting outputs by 1) training models to focus attention on important keyphrases of the story, and 2) promoting generating nongeneric words. We show that the combination of the two leads to more interesting endings.

Pre-trained word vectors are ubiquitous in Natural Language Processing applications. In this paper, we show how training word embeddings jointly with bigram and even trigram embeddings, results in improved unigram embeddings. We claim that training word embeddings along with higher n-gram embeddings helps in the removal of the contextual information from the unigrams, resulting in better stand-alone word embeddings. We empirically show the validity of our hypothesis by outperforming other competing word representation models by a significant margin on a wide variety of tasks. We make our models publicly available.

The recent tremendous success of unsupervised word embeddings in a multitude of applications raises the obvious question if similar methods could be derived to improve embeddings (i.e. semantic representations) of word sequences as well. We present a simple but efficient unsupervised objective to train distributed representations of sentences. Our method outperforms the state-of-the-art unsupervised models on most benchmark tasks, highlighting the robustness of the produced general-purpose sentence embeddings.