Search

Model merging has shown great promise at combining expert models, but the benefit of merging is unclear when merging ``generalist'' models trained on many tasks. We explore merging in the context of large (~100B) models, by recycling checkpoints that exhibit tradeoffs among different tasks. Such checkpoints are often created in the process of developing a frontier model, and many suboptimal ones are usually discarded. Given a pool of model checkpoints obtained from different training runs (e.g., different stages, objectives, hyperparameters, and data mixtures), which naturally show tradeoffs across different language capabilities (e.g., instruction following vs. code generation), we investigate whether merging can recycle such suboptimal models into a Pareto-optimal one. Our optimization algorithm tunes the weight of each checkpoint in a linear combination, resulting in a Pareto-optimal models that outperforms both individual models and merge-based baselines. Further analysis shows that good merges tend to include almost all checkpoints with non-zero weights, indicating that even seemingly bad initial checkpoints can contribute to good final merges., Comment: 13 pages, 9 figures

With the goal of benchmarking generative systems beyond expert software development ability, we introduce Commit0, a benchmark that challenges AI agents to write libraries from scratch. Agents are provided with a specification document outlining the library's API as well as a suite of interactive unit tests, with the goal of producing an implementation of this API accordingly. The implementation is validated through running these unit tests. As a benchmark, Commit0 is designed to move beyond static one-shot code generation towards agents that must process long-form natural language specifications, adapt to multi-stage feedback, and generate code with complex dependencies. Commit0 also offers an interactive environment where models receive static analysis and execution feedback on the code they generate. Our experiments demonstrate that while current agents can pass some unit tests, none can yet fully reproduce full libraries. Results also show that interactive feedback is quite useful for models to generate code that passes more unit tests, validating the benchmarks that facilitate its use.

We report on the nonlinear temporal post-compression of 7 mJ sub-40 fs pulses from a commercial kHz Ti:Sapphire laser down to a record 4 fs duration (1.5 optical cycle) in a compact single-stage gas-filled multipass cell, with 60% overall compression efficiency.

In this paper, we consider contamination by code generation test sets, in particular in their use in modern large language models. We discuss three possible sources of such contamination and show findings supporting each of them: (i) direct data leakage, (ii) indirect data leakage through the use of synthetic data and (iii) overfitting to evaluation sets during model selection. To address this, we release Less Basic Python Problems (LBPP): an uncontaminated new benchmark of 161 prompts with their associated Python solutions. LBPP is released at https://huggingface.co/datasets/CohereForAI/lbpp ., Comment: EMNLP 2024 Findings. 5 main pages, 9 in total

Reward models (RMs) play a critical role in aligning language models through the process of reinforcement learning from human feedback. RMs are trained to predict a score reflecting human preference, which requires significant time and cost for human annotation. Additionally, RMs tend to quickly overfit on superficial features in the training set, hindering their generalization performance on unseen distributions. We propose a novel approach using synthetic natural language critiques generated by large language models to provide additional feedback, evaluating aspects such as instruction following, correctness, and style. This offers richer signals and more robust features for RMs to assess and score on. We demonstrate that high-quality critiques improve the performance and data efficiency of RMs initialized from different pretrained models, reducing the reliance on costly human annotations. Furthermore, incorporating critiques improves both the interpretability and robustness of RM training.

We present an expanded and improved deep-learning (DL) methodology for determining centers of star images on HST/WFPC2 exposures. Previously, we demonstrated that our DL model can eliminate the pixel-phase bias otherwise present in these undersampled images; however that analysis was limited to the central portion of each detector. In the current work we introduce the inclusion of global positions to account for the PSF variation across the entire chip and instrumental magnitudes to account for nonlinear effects such as charge transfer efficiency. The DL model is trained using a unique series of WFPC2 observations of globular cluster 47 Tuc, data sets comprising over 600 dithered exposures taken in each of two filters, F555W and F814W. It is found that the PSF variations across each chip correspond to corrections of the order of 100 mpix, while magnitude effects are at a level of about 10 mpix. Importantly, pixel-phase bias is eliminated with the DL model; whereas, with a classic centering algorithm, the amplitude of this bias can be up to 40 mpix. Our improved DL model yields star-image centers with uncertainties of 8-10 mpix across the full field of view of WFPC2., Comment: accepted for publication by PASP

Humans follow criteria when they execute tasks, and these criteria are directly used to assess the quality of task completion. Therefore, having models learn to use criteria to provide feedback can help humans or models to perform tasks better. However, existing research in this field tends to consider only a limited set of criteria or quality assessment aspects. To fill this gap, we propose a general framework that enables large language models (LLMs) to use comprehensive criteria for a task in delivering natural language feedback on task execution. In particular, we present a model-in-the-loop framework that semi-automatically derives criteria from collected guidelines for different writing tasks and constructs in-context demonstrations for each criterion. We choose three tasks from real-world scenarios to operationalize this idea: paper introduction writing, Python code writing, and Reddit post writing, and evaluate our feedback generation framework using different LLMs. The results reveal the fine-grained effects of incorporating criteria and demonstrations and provide valuable insights on how to teach LLMs to use criteria more effectively., Comment: ACL 2024 findings

AI alignment in the shape of Reinforcement Learning from Human Feedback (RLHF) is increasingly treated as a crucial ingredient for high performance large language models. Proximal Policy Optimization (PPO) has been positioned by recent literature as the canonical method for the RL part of RLHF. However, it involves both high computational cost and sensitive hyperparameter tuning. We posit that most of the motivational principles that led to the development of PPO are less of a practical concern in RLHF and advocate for a less computationally expensive method that preserves and even increases performance. We revisit the formulation of alignment from human preferences in the context of RL. Keeping simplicity as a guiding principle, we show that many components of PPO are unnecessary in an RLHF context and that far simpler REINFORCE-style optimization variants outperform both PPO and newly proposed "RL-free" methods such as DPO and RAFT. Our work suggests that careful adaptation to LLMs alignment characteristics enables benefiting from online RL optimization at low cost., Comment: 27 pages, 7 figures, 2 tables

Extractive summaries are usually presented as lists of sentences with no expected cohesion between them. In this paper, we aim to enforce cohesion whilst controlling for informativeness and redundancy in summaries, in cases where the input exhibits high redundancy. The pipeline controls for redundancy in long inputs as it is consumed, and balances informativeness and cohesion during sentence selection. Our sentence selector simulates human memory to keep track of topics --modeled as lexical chains--, enforcing cohesive ties between noun phrases. Across a variety of domains, our experiments revealed that it is possible to extract highly cohesive summaries that nevertheless read as informative to humans as summaries extracted by only accounting for informativeness or redundancy. The extracted summaries exhibit smooth topic transitions between sentences as signaled by lexical chains, with chains spanning adjacent or near-adjacent sentences.

This paper studies the non-verbal behavior of a conversational agent named Arthur. We propose the development of body movements for this agent, which interacts solely through voice commands, chat, and videos with facial animations. This research aims to analyze users' perceptions regarding the gestures performed by Arthur. This study was conducted with participants who agreed to interact directly or through video with the conversational agent. The main goal is to analyze whether including nonverbal movements alters users' perception so that they feel more comfortable watching the video or interacting in real-time., Comment: Paper presented as Final Project in Computer Science - Pontifical Catholic university of Rio Grande do Sul (Brazil)

We have investigated the properties (e.g., age, metallicity) of the stellar populations of a ring-like tidal stellar stream (or streams) around the edge-on galaxy SPRC047 (z = 0.031) using spectral energy distribution (SED) fits to integrated broad-band aperture flux densities. We used visual images in six different bands and Spitzer/IRAC 3.6 micron data. We have attempted to derive best-fit stellar population parameters (metallicity, age) in three non-contiguous segments of the stream. Due to the very low surface brightness of the stream, we have performed a deconvolution with a Richardson-Lucy type algorithm of the low spatial resolution 3.6 micron IRAC image, thereby reducing the effect of the point-spread-function (PSF) aliased "emission" from the bright edge-on central galaxy at the locations of our three stream segments. Our SED fits that used several different star formation history priors, from an exponentially decaying star formation burst to continuous star formation, indicate that the age-metallicity-dust degeneracy is not resolved, most likely because of inadequate wavelength coverage and low signal-to-noise ratios of the low surface brightness features. We also discuss how future deep visual-near-infrared observations, combined with absolute flux calibration uncertainties at or below the 1 per cent level, complemented by equally well absolute flux calibrated observations in ultraviolet and mid-infrared bands, would improve the accuracy of broad-band SED fitting results for low surface brightness targets, such as stellar streams around nearby galaxies that are not resolved into stars., Comment: Accepted for publication in the Astrophysical Journal

Multi-pass cell (MPC) compressors have proven to be the method of choice for compression of high average power long-pulse Yb lasers. Yet, generating sub-30 fs pulses at high pulse energy with compact and simple components remains a challenge. This work demonstrates an efficient and cost-effective approach for nonlinear pulse compression at high pulse energy using a hybrid air-bulk MPC. By carefully balancing the relative nonlinear contributions of ambient air and fused silica, we achieve strong spectral broadening without dispersion engineering or pressure-control inside the cell at 400 uJ pulse energy. In this way, we compress pulses from 220 fs to 27 fs at 40.3 W of average power (100 kHz repetition rate), enhancing the peak power from 1.6 GW to 10.2 GW while maintaining 78% of the energy within the main pulse. Our approach combines the strengths of gas-filled and bulk compression schemes and exhibits excellent overall optical transmission (91%) and spectral uniformity. Moreover, we utilize the INSIGHT technique to investigate spatio-temporal couplings and geometrical aberrations of the compressed pulse. Our results demonstrate remarkable temporal homogeneity, with an average Strehl ratio of 0.97 consistently observed throughout the entire spectral profile. Additionally, all spectrally-integrated Zernike coefficients for geometrical aberrations maintain values below 0.02{\lambda}.

A Deep Learning (DL) algorithm is built and tested for its ability to determine centers of star images on HST/WFPC2 exposures, in filters F555W and F814W. These archival observations hold great potential for proper-motion studies, but the undersampling in the camera's detectors presents challenges for conventional centering algorithms. Two exquisite data sets of over 600 exposures of the cluster NGC 104 in these filters are used as a testbed for training and evaluation of the DL code. Results indicate a single-measurement standard error of from 8.5 to 11 mpix, depending on detector and filter.This compares favorably to the $\sim20$ mpix achieved with the customary ``effective PSF'' centering procedure for WFPC2 images. Importantly, pixel-phase error is largely eliminated when using the DL method. The current tests are limited to the central portion of each detector; in future studies the DL code will be modified to allow for the known variation of the PSF across the detectors., Comment: accepted for publication by PASP

The BigScience Workshop was a value-driven initiative that spanned one and half years of interdisciplinary research and culminated in the creation of ROOTS, a 1.6TB multilingual dataset that was used to train BLOOM, one of the largest multilingual language models to date. In addition to the technical outcomes and artifacts, the workshop fostered multidisciplinary collaborations around large models, datasets, and their analysis. This in turn led to a wide range of research publications spanning topics from ethics to law, data governance, modeling choices and distributed training. This paper focuses on the collaborative research aspects of BigScience and takes a step back to look at the challenges of large-scale participatory research, with respect to participant diversity and the tasks required to successfully carry out such a project. Our main goal is to share the lessons we learned from this experience, what we could have done better and what we did well. We show how the impact of such a social approach to scientific research goes well beyond the technical artifacts that were the basis of its inception., Comment: Presented at the 2022 NeurIPS Workshop on Broadening Research Collaborations in ML

We present a pilot study to assess the potential of Hyper Suprime-Cam Public Data Release 2 (HSC-PDR2) images for the analysis of extended faint structures within groups of galaxies. We examine the intra-group light (IGL) of the group 400138 ($M_{\rm{dyn}}= 1.3 \pm 0.5 \times 10^{13} $M$_{\odot}$, $z\sim 0.2$) from the Galaxy And Mass Assembly (GAMA) survey using Hyper-Suprime Cam Subaru Strategic Program Public Data Release 2 (HSC-PDR2) images in $g$, $r$, and $i$ bands. We present the most extended IGL measurement to date, reaching down to $\mu_{g}^{\rm{lim}}=30.76$ mag arcsec$^{-2}$ ($3 \sigma$; $10 \times 10$ arcsec$^{2}$) at a semi-major axis of 275 kpc. The IGL shows mean colour values of $g-i=0.92$, $g-r=0.60$, and $r-i=0.32$ ($\pm$0.01). The IGL stellar populations are younger ($2-2.5$ Gyr) and less metal-rich ([Fe/H] $ \sim -$0.4) than those of the host group galaxies. We find a range of IGL fractions as a function of total group luminosity of $\sim 2-36 \%$ depending on the definition of IGL, with larger fractions the bluer the observation wavelength. The early-type to late-type galaxy ratio suggests that 400138 is a more evolved group, dominated by ETGs, and the IGL fraction agrees with that of other similarly evolved groups. These results are consistent with tidal stripping of the outer parts of Milky Way-like galaxies as the main driver of the IGL build-up. This is supported by the detection of substructure in the IGL towards the galaxy member 1660615 suggesting a recent interaction ($<1$ Gyr ago) of that galaxy with the core of the group., Comment: 19 pages, 8 figures, 2 tables, accepted for publication in MNRAS

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

Extractive summaries are usually presented as lists of sentences with no expected cohesion between them and with plenty of redundant information if not accounted for. In this paper, we investigate the trade-offs incurred when aiming to control for inter-sentential cohesion and redundancy in extracted summaries, and their impact on their informativeness. As case study, we focus on the summarization of long, highly redundant documents and consider two optimization scenarios, reward-guided and with no supervision. In the reward-guided scenario, we compare systems that control for redundancy and cohesion during sentence scoring. In the unsupervised scenario, we introduce two systems that aim to control all three properties -- informativeness, redundancy, and cohesion -- in a principled way. Both systems implement a psycholinguistic theory that simulates how humans keep track of relevant content units and how cohesion and non-redundancy constraints are applied in short-term memory during reading. Extensive automatic and human evaluations reveal that systems optimizing for -- among other properties -- cohesion are capable of better organizing content in summaries compared to systems that optimize only for redundancy, while maintaining comparable informativeness. We find that the proposed unsupervised systems manage to extract highly cohesive summaries across varying levels of document redundancy, although sacrificing informativeness in the process. Finally, we lay evidence as to how simulated cognitive processes impact the trade-off between the analyzed summary properties., Comment: Accepted to JAIR

What are the units of text that we want to model? From bytes to multi-word expressions, text can be analyzed and generated at many granularities. Until recently, most natural language processing (NLP) models operated over words, treating those as discrete and atomic tokens, but starting with byte-pair encoding (BPE), subword-based approaches have become dominant in many areas, enabling small vocabularies while still allowing for fast inference. Is the end of the road character-level model or byte-level processing? In this survey, we connect several lines of work from the pre-neural and neural era, by showing how hybrid approaches of words and characters as well as subword-based approaches based on learned segmentation have been proposed and evaluated. We conclude that there is and likely will never be a silver bullet singular solution for all applications and that thinking seriously about tokenization remains important for many applications., Comment: 15 page preprint

Softmax is the de facto standard in modern neural networks for language processing when it comes to normalizing logits. However, by producing a dense probability distribution each token in the vocabulary has a nonzero chance of being selected at each generation step, leading to a variety of reported problems in text generation. $\alpha$-entmax of Peters et al. (2019, arXiv:1905.05702) solves this problem, but is considerably slower than softmax. In this paper, we propose an alternative to $\alpha$-entmax, which keeps its virtuous characteristics, but is as fast as optimized softmax and achieves on par or better performance in machine translation task., Comment: Findings of NAACL 2022

The power of natural language generation models has provoked a flurry of interest in automatic methods to detect if a piece of text is human or machine-authored. The problem so far has been framed in a standard supervised way and consists in training a classifier on annotated data to predict the origin of one given new document. In this paper, we frame the problem in an unsupervised and distributional way: we assume that we have access to a large collection of unannotated documents, a big fraction of which is machine-generated. We propose a method to detect those machine-generated documents leveraging repeated higher-order n-grams, which we show over-appear in machine-generated text as compared to human ones. That weak signal is the starting point of a self-training setting where pseudo-labelled documents are used to train an ensemble of classifiers. Our experiments show that leveraging that signal allows us to rank suspicious documents accurately. Precision at 5000 is over 90% for top-k sampling strategies, and over 80% for nucleus sampling for the largest model we used (GPT2-large). The drop with increased size of model is small, which could indicate that the results hold for other current and future large language models., Comment: 10 pages

We consider the problem of multilingual unsupervised machine translation, translating to and from languages that only have monolingual data by using auxiliary parallel language pairs. For this problem the standard procedure so far to leverage the monolingual data is back-translation, which is computationally costly and hard to tune. In this paper we propose instead to use denoising adapters, adapter layers with a denoising objective, on top of pre-trained mBART-50. In addition to the modularity and flexibility of such an approach we show that the resulting translations are on-par with back-translating as measured by BLEU, and furthermore it allows adding unseen languages incrementally., Comment: Accepted as a long paper to EMNLP 2021

The present paper aims at analyzing the topological content of the complex trajectories that weeder-autonomous robots follow in operation. We will prove that the topological descriptors of these trajectories are affected by the robot environment as well as by the robot state, with respect to maintenance operations. Topological Data Analysis will be used for extracting the trajectory descriptors, based on homology persistence. Then, appropriate metrics will be applied in order to compare that topological representation of the trajectories, for classifying them or for making efficient pattern recognition.

Social media services such as TripAdvisor and Foursquare can provide opportunities for users to exchange their opinions about urban green space (UGS). Visitors can exchange their experiences with parks, woods, and wetlands in social communities via social networks. In this work, we implement a unified topic modeling approach to reveal UGS characteristics. Leveraging Artificial Intelligence techniques for opinion mining using the mentioned platforms (e.g., TripAdvisor and Foursquare) reviews is a novel application to UGS quality assessments. We show how specific characteristics of different green spaces can be explored by using a tailor-optimized sentiment analysis model. Such an application can support local authorities and stakeholders in understanding--and justification for--future urban green space investments., Comment: 6 pages, 4 figures

As neural machine translation (NMT) systems become an important part of professional translator pipelines, a growing body of work focuses on combining NMT with terminologies. In many scenarios and particularly in cases of domain adaptation, one expects the MT output to adhere to the constraints provided by a terminology. In this work, we propose metrics to measure the consistency of MT output with regards to a domain terminology. We perform studies on the COVID-19 domain over 5 languages, also performing terminology-targeted human evaluation. We open-source the code for computing all proposed metrics: https://github.com/mahfuzibnalam/terminology_evaluation, Comment: preprint

Summarization systems face the core challenge of identifying and selecting important information. In this paper, we tackle the problem of content selection in unsupervised extractive summarization of long, structured documents. We introduce a wide range of heuristics that leverage cognitive representations of content units and how these are retained or forgotten in human memory. We find that properties of these representations of human memory can be exploited to capture relevance of content units in scientific articles. Experiments show that our proposed heuristics are effective at leveraging cognitive structures and the organization of the document (i.e.\ sections of an article), and automatic and human evaluations provide strong evidence that these heuristics extract more summary-worthy content units.

There is an ongoing debate in the NLP community whether modern language models contain linguistic knowledge, recovered through so-called probes. In this paper, we study whether linguistic knowledge is a necessary condition for the good performance of modern language models, which we call the \textit{rediscovery hypothesis}. In the first place, we show that language models that are significantly compressed but perform well on their pretraining objectives retain good scores when probed for linguistic structures. This result supports the rediscovery hypothesis and leads to the second contribution of our paper: an information-theoretic framework that relates language modeling objectives with linguistic information. This framework also provides a metric to measure the impact of linguistic information on the word prediction task. We reinforce our analytical results with various experiments, both on synthetic and on real NLP tasks in English.

Continued population growth and urbanization is shifting research to consider the quality of urban green space over the quantity of these parks, woods, and wetlands. The quality of urban green space has been hitherto measured by expert assessments, including in-situ observations, surveys, and remote sensing analyses. Location data platforms, such as TripAdvisor, can provide people's opinion on many destinations and experiences, including UGS. This paper leverages Artificial Intelligence techniques for opinion mining and text classification using such platform's reviews as a novel approach to urban green space quality assessments. Natural Language Processing is used to analyze contextual information given supervised scores of words by implementing computational analysis. Such an application can support local authorities and stakeholders in their understanding of and justification for future investments in urban green space.

It is standard procedure these days to solve Information Extraction task by fine-tuning large pre-trained language models. This is not the case for generation task, which relies on a variety of techniques for controlled language generation. In this paper, we describe a system that fine-tunes a natural language generation model for the problem of solving Writer's Block. The fine-tuning changes the conditioning to also include the right context in addition to the left context, as well as an optional list of entities, the size, the genre and a summary of the paragraph that the human author wishes to generate. Our proposed fine-tuning obtains excellent results, even with a small number of epochs and a total cost of USD 150. The system can be accessed as a web-service, and all the code is released. A video showcasing the interface and the model is also available., Comment: Accepted at EACL 2021

We release a multilingual neural machine translation model, which can be used to translate text in the biomedical domain. The model can translate from 5 languages (French, German, Italian, Korean and Spanish) into English. It is trained with large amounts of generic and biomedical data, using domain tags. Our benchmarks show that it performs near state-of-the-art both on news (generic domain) and biomedical test sets, and that it outperforms the existing publicly released models. We believe that this release will help the large-scale multilingual analysis of the digital content of the COVID-19 crisis and of its effects on society, economy, and healthcare policies. We also release a test set of biomedical text for Korean-English. It consists of 758 sentences from official guidelines and recent papers, all about COVID-19., Comment: https://github.com/naver/covid19-nmt

Organic light-emitting diodes (OLEDs) have successfully entered the display market and continue to be attractive for many other applications. As state-of-the-art OLEDs can reach an internal quantum efficiency (IQE) of almost 100 %, light outcoupling remains one of the major screws left to be turned. The fact that no superior outcoupling structure has been found underlines that further investigations are needed to understand their prospect. In this paper, we use two-dimensional titanium dioxide (2D TiO$_2$) block arrays as a model of an internal light outcoupling structure and investigate the influence of its geometrical parameters on achieving the highest external quantum efficiency (EQE) for OLEDs. The multivariable problem is evaluated with the visual assistance of scatter plots, which enables us to propose an optimal period range and block width-to-distance ratio. The highest EQE achieved is 45.2 % with internal and external structures. This work contributes to the highly desired prediction of ideal light outcoupling structures in the future., Comment: 14 pages, 8 Postscript figures

Analysis pipelines commonly use high-level technologies that are popular when created, but are unlikely to be readable, executable, or sustainable in the long term. A set of criteria is introduced to address this problem: Completeness (no execution requirement beyond a minimal Unix-like operating system, no administrator privileges, no network connection, and storage primarily in plain text); modular design; minimal complexity; scalability; verifiable inputs and outputs; version control; linking analysis with narrative; and free and open source software. As a proof of concept, we introduce "Maneage" (Managing data lineage), enabling cheap archiving, provenance extraction, and peer verification that has been tested in several research publications. We show that longevity is a realistic requirement that does not sacrifice immediate or short-term reproducibility. The caveats (with proposed solutions) are then discussed and we conclude with the benefits for the various stakeholders. This article is itself a Maneage'd project (project commit 54e4eb2)., Comment: Published version. The downloadable source (on arXiv) includes the full/automatic reproduction resources (scripts, config files and input data links). Git repository: https://git.maneage.org/paper-concept.git (also on Software Heritage), Zenodo: https://doi.org/10.5281/zenodo.3872247

We address the problem of unsupervised abstractive summarization of collections of user generated reviews with self-supervision and control. We propose a self-supervised setup that considers an individual document as a target summary for a set of similar documents. This setting makes training simpler than previous approaches by relying only on standard log-likelihood loss. We address the problem of hallucinations through the use of control codes, to steer the generation towards more coherent and relevant summaries.Finally, we extend the Transformer architecture to allow for multiple reviews as input. Our benchmarks on two datasets against graph-based and recent neural abstractive unsupervised models show that our proposed method generates summaries with a superior quality and relevance.This is confirmed in our human evaluation which focuses explicitly on the faithfulness of generated summaries We also provide an ablation study, which shows the importance of the control setup in controlling hallucinations and achieve high sentiment and topic alignment of the summaries with the input reviews., Comment: 18 pages including 5 pages appendix