Perils of Self-Feedback: Self-Bias Amplifies in Large Language Models

Recent studies show that self-feedback improves large language models (LLMs) on certain tasks while worsens other tasks. We discovered that such a contrary is due to LLM's bias towards their own output. In this paper, we formally define LLM's self-bias -- the tendency to favor its own generation -- using two statistics. We analyze six LLMs on translation, constrained text generation, and mathematical reasoning tasks. We find that self-bias is prevalent in all examined LLMs across multiple languages and tasks. Our analysis reveals that while the self-refine pipeline improves the fluency and understandability of model outputs, it further amplifies self-bias. To mitigate such biases, we discover that larger model size and external feedback with accurate assessment can significantly reduce bias in the self-refine pipeline, leading to actual performance improvement in downstream tasks.

Pinpoint, Not Criticize: Refining Large Language Models via Fine-Grained Actionable Feedback

Recent improvements in text generation have leveraged human feedback to improve the quality of the generated output. However, human feedback is not always available, especially during inference. In this work, we propose an inference time optimization method FITO to use fine-grained actionable feedback in the form of error type, error location and severity level that are predicted by a learned error pinpoint model for iterative refinement. FITO starts with an initial output, then iteratively incorporates the feedback via a refinement model that generates an improved output conditioned on the feedback. Given the uncertainty of consistent refined samples at iterative steps, we formulate iterative refinement into a local search problem and develop a simulated annealing based algorithm that balances exploration of the search space and optimization for output quality. We conduct experiments on three text generation tasks, including machine translation, long-form question answering (QA) and topical summarization. We observe 0.8 and 0.7 MetricX gain on Chinese-English and English-German translation, 4.5 and 1.8 ROUGE-L gain at long form QA and topic summarization respectively, with a single iteration of refinement. With our simulated annealing algorithm, we see further quality improvements, including up to 1.7 MetricX improvements over the baseline approach.

Automatically Correcting Large Language Models: Surveying the landscape of diverse self-correction strategies

Large language models (LLMs) have demonstrated remarkable performance across a wide array of NLP tasks. However, their efficacy is undermined by undesired and inconsistent behaviors, including hallucination, unfaithful reasoning, and toxic content. A promising approach to rectify these flaws is self-correction, where the LLM itself is prompted or guided to fix problems in its own output. Techniques leveraging automated feedback -- either produced by the LLM itself or some external system -- are of particular interest as they are a promising way to make LLM-based solutions more practical and deployable with minimal human feedback. This paper presents a comprehensive review of this emerging class of techniques. We analyze and taxonomize a wide array of recent work utilizing these strategies, including training-time, generation-time, and post-hoc correction. We also summarize the major applications of this strategy and conclude by discussing future directions and challenges.

INSTRUCTSCORE: Towards Explainable Text Generation Evaluation with Automatic Feedback

The field of automatic evaluation of text generation made tremendous progress in the last few years. In particular, since the advent of neural metrics, like COMET, BLEURT, and SEScore2, the newest generation of metrics show a high correlation with human judgment. Unfortunately, quality scores generated with neural metrics are not interpretable, and it is unclear which part of the generation output is criticized by the metrics. To address this limitation, we present INSTRUCTSCORE, an open-source, explainable evaluation metric for text generation. By harnessing both explicit human instruction and the implicit knowledge of GPT4, we fine-tune a LLAMA model to create an evaluative metric that can produce a diagnostic report aligned with human judgment. We evaluate INSTRUCTSCORE on the WMT22 Zh-En translation task, where our 7B model surpasses other LLM-based baselines, including those based on 175B GPT3. Impressively, our INSTRUCTSCORE, even without direct supervision from human-rated data, achieves performance levels on par with state-of-the-art metrics like COMET22, which was fine-tuned on human ratings.

CausalDialogue: Modeling Utterance-level Causality in Conversations

Despite their widespread adoption, neural conversation models have yet to exhibit natural chat capabilities with humans. In this research, we examine user utterances as causes and generated responses as effects, recognizing that changes in a cause should produce a different effect. To further explore this concept, we have compiled and expanded upon a new dataset called CausalDialogue through crowd-sourcing. This dataset includes multiple cause-effect pairs within a directed acyclic graph (DAG) structure. Our analysis reveals that traditional loss functions can struggle to effectively incorporate the DAG structure, leading us to propose a causality-enhanced method called Exponential Maximum Average Treatment Effect (ExMATE) to enhance the impact of causality at the utterance level in training neural conversation models. To evaluate the effectiveness of this approach, we have built a comprehensive benchmark using the CausalDialogue dataset leveraging large-scale pre-trained language models, and have assessed the results through both human and automatic evaluation metrics for coherence, diversity, and agility. Our findings show that current techniques are still unable to effectively address conversational DAGs, and that the ExMATE method can improve the diversity and agility of conventional loss functions while maintaining coherence.

SEScore2: Retrieval Augmented Pretraining for Text Generation Evaluation

Is it possible to leverage large scale raw and raw parallel corpora to build a general learned metric? Existing learned metrics have gaps to human judgements, are model-dependent or are limited to the domains or tasks where human ratings are available. In this paper, we propose SEScore2, a model-based metric pretrained over million-scale synthetic dataset constructed by our novel retrieval augmented data synthesis pipeline. SEScore2 achieves high correlation to human judgements without any human rating supervisions. Importantly, our unsupervised SEScore2 can outperform supervised metrics, which are trained on the News human ratings, at the TED domain. We evaluate SEScore2 over four text generation tasks across three languages. SEScore2 outperforms all prior unsupervised evaluation metrics in machine translation, speech translation, data-to-text and dialogue generation, with average Kendall improvements 0.158. SEScore2 even outperforms SOTA supervised BLEURT at data-to-text, dialogue generation and overall correlation.

Not All Errors are Equal: Learning Text Generation Metrics using Stratified Error Synthesis

Is it possible to build a general and automatic natural language generation (NLG) evaluation metric? Existing learned metrics either perform unsatisfactorily or are restricted to tasks where large human rating data is already available. We introduce SESCORE, a model-based metric that is highly correlated with human judgements without requiring human annotation, by utilizing a novel, iterative error synthesis and severity scoring pipeline. This pipeline applies a series of plausible errors to raw text and assigns severity labels by simulating human judgements with entailment. We evaluate SESCORE against existing metrics by comparing how their scores correlate with human ratings. SESCORE outperforms all prior unsupervised metrics on multiple diverse NLG tasks including machine translation, image captioning, and WebNLG text generation. For WMT 20/21 En-De and Zh-En, SESCORE improve the average Kendall correlation with human judgement from 0.154 to 0.195. SESCORE even achieves comparable performance to the best supervised metric COMET, despite receiving no human-annotated training data.

Visualize Before You Write: Imagination-Guided Open-Ended Text Generation

Recent advances in text-to-image synthesis make it possible to visualize machine imaginations for a given context. On the other hand, when generating text, human writers are gifted at creative visualization, which enhances their writings by forming imaginations as blueprints before putting down the stories in words. Inspired by such a cognitive process, we ask the natural question of whether we can endow machines with the same ability to utilize visual information and construct a general picture of the context to guide text generation. In this work, we propose iNLG that uses machine-generated images to guide language models (LM) in open-ended text generation. The experiments and analyses demonstrate the effectiveness of iNLG on open-ended text generation tasks, including text completion, story generation, and concept-to-text generation in few-shot scenarios. Both automatic metrics and human evaluations verify that the text snippets generated by our iNLG are coherent and informative while displaying minor degeneration.

Anticipating the Unseen Discrepancy for Vision and Language Navigation

Vision-Language Navigation requires the agent to follow natural language instructions to reach a specific target. The large discrepancy between seen and unseen environments makes it challenging for the agent to generalize well. Previous studies propose data augmentation methods to mitigate the data bias explicitly or implicitly and provide improvements in generalization. However, they try to memorize augmented trajectories and ignore the distribution shifts under unseen environments at test time. In this paper, we propose an Unseen Discrepancy Anticipating Vision and Language Navigation (DAVIS) that learns to generalize to unseen environments via encouraging test-time visual consistency. Specifically, we devise: 1) a semi-supervised framework DAVIS that leverages visual consistency signals across similar semantic observations. 2) a two-stage learning procedure that encourages adaptation to test-time distribution. The framework enhances the basic mixture of imitation and reinforcement learning with Momentum Contrast to encourage stable decision-making on similar observations under a joint training stage and a test-time adaptation stage. Extensive experiments show that DAVIS achieves model-agnostic improvement over previous state-of-the-art VLN baselines on R2R and RxR benchmarks. Our source code and data are in supplemental materials.

Neuro-Symbolic Causal Language Planning with Commonsense Prompting

Language planning aims to implement complex high-level goals by decomposition into sequential simpler low-level steps. Such procedural reasoning ability is essential for applications such as household robots and virtual assistants. Although language planning is a basic skill set for humans in daily life, it remains a challenge for large language models (LLMs) that lack deep-level commonsense knowledge in the real world. Previous methods require either manual exemplars or annotated programs to acquire such ability from LLMs. In contrast, this paper proposes Neuro-Symbolic Causal Language Planner (CLAP) that elicits procedural knowledge from the LLMs with commonsense-infused prompting. Pre-trained knowledge in LLMs is essentially an unobserved confounder that causes spurious correlations between tasks and action plans. Through the lens of a Structural Causal Model (SCM), we propose an effective strategy in CLAP to construct prompts as a causal intervention toward our SCM. Using graph sampling techniques and symbolic program executors, our strategy formalizes the structured causal prompts from commonsense knowledge bases. CLAP obtains state-of-the-art performance on WikiHow and RobotHow, achieving a relative improvement of 5.28% in human evaluations under the counterfactual setting. This indicates the superiority of CLAP in causal language planning semantically and sequentially.