In this work we face a challenging puzzle video game: A Good Snowman is Hard to Build. The objective of the game is to build snowmen by moving and stacking snowballs on a discrete grid. For the sake of player engagement with the game, it is interesting to avoid that a player finds a much easier solution than the one the designer expected. Therefore, having tools that are able to certify the optimality of solutions is crucial. Although the game can be stated as a planning problem and can be naturally modelled in PDDL, we show that a direct translation to SAT clearly outperforms off-the-shelf state-of-the-art planners. As we show, this is mainly due to the fact that reachability properties can be easily modelled in SAT, allowing for shorter plans, whereas using axioms to express a reachability derived predicate in PDDL does not result in any significant reduction of solving time with the considered planners. We deal with a set of 51 levels, both original and crafted, solving 43 and with 8 challenging instances still remaining to be solved.

Many puzzle video games, like Sokoban, involve moving some agent in a maze. The reachable locations are usually apparent for a human player, and the difficulty of the game is mainly related to performing actions on objects, such as pushing (reachable) boxes. For this reason, the difficulty of a particular level is often measured as the number of actions on objects, other than agent walking, needed to find a solution. In this paper we study CP and SAT approaches for solving these kind of problems. We review some reachability encodings and propose a new one. We empirically show that the new encoding is well-suited for solving puzzle problems in the planning as SAT paradigm, especially when considering the execution of several actions in parallel.

Reinforcement learning from human feedback (RLHF) has become a pivot technique in aligning large language models (LLMs) with human preferences. In RLHF practice, preference data plays a crucial role in bridging human proclivity and LLMs. However, the scarcity of diverse, naturalistic datasets of human preferences on LLM outputs at scale poses a great challenge to RLHF as well as feedback learning research within the open-source community. Current preference datasets, either proprietary or limited in size and prompt variety, result in limited RLHF adoption in open-source models and hinder further exploration. We meticulously devise annotation instructions and employ GPT-4 to offer detailed feedback in both numerical and textual forms. Experimental results indicate that our models outperform existing open-source models, achieving top performance across multiple benchmarks. Large language models (LLMs), represented by ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI, 2023), have demonstrated proficiency in generating fluent text as well as solving various languageoriented tasks. Trained on massive corpora through likelihood maximization techniques, these LLMs have exhibited remarkable generalization and equipped the ability to execute diverse tasks in response to user directives (Ouyang et al., 2022; Wei et al., 2022; Sanh et al., 2022). Unfortunately, relying solely on likelihood maximization during training leads to well-known issues - LLMs may generate convincing but incorrect or unsafe content that deviates from human preferences (Stiennon et al., 2020; Ouyang et al., 2022; Perez et al., 2022). To further align LLMs with human preferences, reinforcement learning from human feedback (RLHF) (Ouyang et al., 2022; Askell et al., 2021; Bai et al., 2022a; Touvron et al., 2023b) has been introduced and widely adopted by leading corporations. RLHF builds upon preference data, which rates and compares different responses given the same prompt. Typically, RLHF trains a reward model on preference data and then applies RL algorithms such as Proximal Policy Optimization (PPO) (Schulman et al., 2017) on LLMs to optimize the rewards (OpenAI, 2022; 2023; Touvron et al., 2023b; Bai et al., 2022a). While proprietary models have largely capitalized on RLHF's potential to produce outputs that are both more useful and safer, a significant gap persists in the open-source community. As a result, few open-source models adopt RLHF as it demonstrates marginal gains, which critically hinders successful RLHF practice and further research.

It has become mainstream in computer vision and other machine learning domains to reuse backbone networks pre-trained on large datasets as preprocessors. Typically, the last layer is replaced by a shallow learning machine of sorts; the newly-added classification head and (optionally) deeper layers are fine-tuned on a new task. Due to its strong performance and simplicity, a common pre-trained backbone network is ResNet152.However, ResNet152 is relatively large and induces inference latency. In many cases, a compact and efficient backbone with similar performance would be preferable over a larger, slower one. This paper investigates techniques to reuse a pre-trained backbone with the objective of creating a smaller and faster model. Starting from a large ResNet152 backbone pre-trained on ImageNet, we first reduce it from 51 blocks to 5 blocks, reducing its number of parameters and FLOPs by more than 6 times, without significant performance degradation. Then, we split the model after 3 blocks into several branches, while preserving the same number of parameters and FLOPs, to create an ensemble of sub-networks to improve performance. Our experiments on a large benchmark of $40$ image classification datasets from various domains suggest that our techniques match the performance (if not better) of ``classical backbone fine-tuning'' while achieving a smaller model size and faster inference speed.

Reward modeling (a.k.a., preference modeling) is instrumental for aligning large language models with human preferences, particularly within the context of reinforcement learning from human feedback (RLHF). While conventional reward models (RMs) have exhibited remarkable scalability, they oft struggle with fundamental functionality such as arithmetic computation, code execution, and factual lookup. In this paper, we propose a tool-augmented preference modeling approach, named \name, to address these limitations by empowering RMs with access to external environments, including calculators and search engines. This approach not only fosters synergy between tool utilization and reward grading but also enhances interpretive capacity and scoring reliability. Our study delves into the integration of external tools into RMs, enabling them to interact with diverse external sources and construct task-specific tool engagement and reasoning traces in an autoregressive manner. We validate our approach across a wide range of domains, incorporating seven distinct external tools. Our experimental results demonstrate a noteworthy overall improvement of 17.7% across eight tasks in preference ranking. Furthermore, our approach outperforms Gopher 280B by 7.3% on TruthfulQA task in zero-shot evaluation. In human evaluations, RLHF trained with Themis attains an average win rate of 32% when compared to baselines across four distinct tasks. Additionally, we provide a comprehensive collection of tool-related RM datasets, incorporating data from seven distinct tool APIs, totaling 15,000 instances. We anticipate that this publicly available dataset will facilitate and inspire further research advancements in the field.

Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts that align with human texts across different aspects. In particular, sampling-based methods produce less-repetitive texts which are often disjunctive in discourse, while search-based methods maintain topic coherence at the cost of increased repetition. Overall, these methods fall short in achieving holistic alignment across a broad range of aspects. In this work, we frame decoding from a language model as an optimization problem with the goal of strictly matching the expected performance with human texts measured by multiple metrics of desired aspects simultaneously. The resulting decoding distribution enjoys an analytical solution that scales the input language model distribution via a sequence-level energy function defined by these metrics. And most importantly, we prove that this induced distribution is guaranteed to improve the perplexity on human texts, which suggests a better approximation to the underlying distribution of human texts. To facilitate tractable sampling from this globally normalized distribution, we adopt the Sampling-Importance-Resampling technique. Experiments on various domains and model scales demonstrate the superiority of our method in metrics alignment with human texts and human evaluation over strong baselines.

Structured documents analysis and recognition are essential for modern online on-boarding processes, and document localization is a crucial step to achieve reliable key information extraction. While deep-learning has become the standard technique used to solve document analysis problems, real-world applications in industry still face the limited availability of labelled data and of computational resources when training or fine-tuning deep-learning models. To tackle these challenges, we propose SDL-Net: a novel U-Net like encoder-decoder architecture for the localization of structured documents. Our approach allows pre-training the encoder of SDL-Net on a generic dataset containing samples of various document classes, and enables fast and data-efficient fine-tuning of decoders to support the localization of new document classes. We conduct extensive experiments on a proprietary dataset of structured document images to demonstrate the effectiveness and the generalization capabilities of the proposed approach.

Large language models (LLMs) often encounter knowledge conflicts, scenarios where discrepancy arises between the internal parametric knowledge of LLMs and non-parametric information provided in the prompt context. In this work we ask what are the desiderata for LLMs when a knowledge conflict arises and whether existing LLMs fulfill them. We posit that LLMs should 1) identify knowledge conflicts, 2) pinpoint conflicting information segments, and 3) provide distinct answers or viewpoints in conflicting scenarios. To this end, we introduce KNOWLEDGE CONFLICT, an evaluation framework for simulating contextual knowledge conflicts and quantitatively evaluating to what extent LLMs achieve these goals. KNOWLEDGE CONFLICT includes diverse and complex situations of knowledge conflict, knowledge from diverse entities and domains, two synthetic conflict creation methods, and settings with progressively increasing difficulty to reflect realistic knowledge conflicts. Extensive experiments with the KNOWLEDGE CONFLICT framework reveal that while LLMs perform well in identifying the existence of knowledge conflicts, they struggle to determine the specific conflicting knowledge and produce a response with distinct answers amidst conflicting information. To address these challenges, we propose new instruction-based approaches that augment LLMs to better achieve the three goals. Further analysis shows that abilities to tackle knowledge conflicts are greatly impacted by factors such as knowledge domain and prompt text, while generating robust responses to knowledge conflict scenarios remains an open research question.

Speech enhancement aims to improve the quality of speech signals in terms of quality and intelligibility, and speech editing refers to the process of editing the speech according to specific user needs. In this paper, we propose a Unified Speech Enhancement and Editing (uSee) model with conditional diffusion models to handle various tasks at the same time in a generative manner. Specifically, by providing multiple types of conditions including self-supervised learning embeddings and proper text prompts to the score-based diffusion model, we can enable controllable generation of the unified speech enhancement and editing model to perform corresponding actions on the source speech. Our experiments show that our proposed uSee model can achieve superior performance in both speech denoising and dereverberation compared to other related generative speech enhancement models, and can perform speech editing given desired environmental sound text description, signal-to-noise ratios (SNR), and room impulse responses (RIR). Demos of the generated speech are available at https://muqiaoy.github.io/usee.

Healthcare data is often split into medium/small-sized collections across multiple hospitals and access to it is encumbered by privacy regulations. This brings difficulties to use them for the development of machine learning and deep learning models, which are known to be data-hungry. One way to overcome this limitation is to use collaborative learning (CL) methods, which allow hospitals to work collaboratively to solve a task, without the need to explicitly share local data. In this paper, we address a prostate segmentation problem from MRI in a collaborative scenario by comparing two different approaches: federated learning (FL) and consensus-based methods (CBM). To the best of our knowledge, this is the first work in which CBM, such as label fusion techniques, are used to solve a problem of collaborative learning. In this setting, CBM combine predictions from locally trained models to obtain a federated strong learner with ideally improved robustness and predictive variance properties. Our experiments show that, in the considered practical scenario, CBMs provide equal or better results than FL, while being highly cost-effective. Our results demonstrate that the consensus paradigm may represent a valid alternative to FL for typical training tasks in medical imaging.