Robotic manipulation, owing to its multi-modal nature, often faces significant training ambiguity, necessitating explicit instructions to clearly delineate the manipulation details in tasks. In this work, we highlight that vision instruction is naturally more comprehensible to recent robotic policies than the commonly adopted text instruction, as these policies are born with some vision understanding ability like human infants. Building on this premise and drawing inspiration from cognitive science, we introduce the robotic imagery paradigm, which realizes large-scale robotic data pre-training without text annotations. Additionally, we propose the robotic gaze strategy that emulates the human eye gaze mechanism, thereby guiding subsequent actions and focusing the attention of the policy on the manipulated object. Leveraging these innovations, we develop VIRT, a fully Transformer-based policy. We design comprehensive tasks using both a physical robot and simulated environments to assess the efficacy of VIRT. The results indicate that VIRT can complete very competitive tasks like "opening the lid of a tightly sealed bottle", and the proposed techniques boost the success rates of the baseline policy on diverse challenging tasks from nearly 0% to more than 65%. The key insight that supports this work is existing robotic policies are akin to human infants, who are born with visual perception and reasoning abilities but do not comprehend natural language according to previous cognitive science literature (Colombo & Mitchell, 2009). Specifically, visual signal serves as the primary information source of recent robotic policies, and the backbones of these policies are pre-trained with large-scale image datasets before the robotic data based training (He et al., 2016; Oquab et al., 2024). Therefore, the policies begin with a basic visual understanding capability like human infants. By contrast, natural language inputs are rarely incorporated into the process of pre-training these backbones, suggesting the lack of natural language knowledge in these policies.

Vision-Language Models (VLMs) are powerful tools for processing and understanding text and images. We study the processing of visual tokens in the language model component of LLaVA, a prominent VLM. Our approach focuses on analyzing the localization of object information, the evolution of visual token representations across layers, and the mechanism of integrating visual information for predictions. Through ablation studies, we demonstrated that object identification accuracy drops by over 70% when object-specific tokens are removed. We observed that visual token representations become increasingly interpretable in the vocabulary space across layers, suggesting an alignment with textual tokens corresponding to image content. Finally, we found that the model extracts object information from these refined representations at the last token position for prediction, mirroring the process in text-only language models for factual association tasks. These findings provide crucial insights into how VLMs process and integrate visual information, bridging the gap between our understanding of language and vision models, and paving the way for more interpretable and controllable multimodal systems. Vision-Language Models (VLMs) take an image and text as input, and generate a text output. They have become powerful tools in processing and understanding text and images, enabling a wide range of applications from question answering to image captioning (Liu et al., 2023b; Li et al., 2023; Alayrac et al., 2022).

One essential advantage of recurrent neural networks (RNNs) over transformerbased language models is their linear computational complexity concerning the sequence length, which makes them much faster in handling long sequences during inference. However, most publicly available RNNs (e.g., Mamba and RWKV) are trained on sequences with less than 10K tokens, and their effectiveness in longer contexts remains largely unsatisfying so far. In this paper, we study the cause of the inability to process long context for RNNs and suggest critical mitigations. We examine two practical concerns when applying state-of-the-art RNNs to long contexts: (1) the inability to extrapolate to inputs longer than the training length and (2) the upper bound of memory capacity. Addressing the first concern, we first investigate state collapse (SC), a phenomenon that causes severe performance degradation on sequence lengths not encountered during training. With controlled experiments, we attribute this to overfitting due to the recurrent state being overparameterized for the training length. For the second concern, we train a series of Mamba-2 models on long documents to empirically estimate the recurrent state capacity in language modeling and passkey retrieval. Then, three SC mitigation methods are proposed to improve Mamba-2's length generalizability, allowing the model to process more than 1M tokens without SC. We also find that the recurrent state capacity in passkey retrieval scales exponentially to the state size, and we empirically train a Mamba-2 370M with near-perfect passkey retrieval accuracy on 256K context length. This suggests a promising future for RNN-based long-context modeling. Recent transformer-based language models (Vaswani et al., 2017; Brown et al., 2020; Touvron et al., 2023; Achiam et al., 2023; Dubey et al., 2024) have demonstrated impressive capabilities in reasoning over long sequences with thousands and even millions of tokens (Team, 2024a;b; GLM et al., 2024).

Automatic LLM benchmarks, such as AlpacaEval 2.0, Arena-Hard-Auto, and MT-Bench, have become popular for evaluating language models due to their cost-effectiveness and scalability compared to human evaluation. Achieving high win rates on these benchmarks can significantly boost the promotional impact of newly released language models. This promotional benefit may motivate tricks, such as manipulating model output length or style to game win rates, even though several mechanisms have been developed to control length and disentangle style to reduce gameability. Nonetheless, we show that even a "null model" that always outputs a constant response (irrelevant to input instructions) can cheat automatic benchmarks and achieve top-ranked win rates: an 86.5% LC win rate on AlpacaEval 2.0; an 83.0 score on Arena-Hard-Auto; and a 9.55 score on MT-Bench. Moreover, the crafted cheating outputs are transferable because we assume that the instructions of these benchmarks (e.g., 805 samples of AlpacaEval 2.0) are private and cannot be accessed. While our experiments are primarily proof-of-concept, an adversary could use LLMs to generate more imperceptible cheating responses, unethically benefiting from high win rates and promotional impact. Our findings call for the development of anti-cheating mechanisms for reliable automatic benchmarks. The code is available at https://github.com/sail-sg/Cheating-LLM-Benchmarks.

The use of small language models (SLMs), herein defined as models with less than three billion parameters, is increasing across various domains and applications. Due to their ability to run on more accessible hardware and preserve user privacy, SLMs possess the potential to democratize access to language models for individuals of different socioeconomic status and with different privacy preferences. This study assesses several state-of-the-art SLMs (e.g., Apple's OpenELM, Microsoft's Phi, Google's Gemma, and the Tinyllama project) for use in the financial domain to support the development of financial literacy LMs. Democratizing access to quality financial information for those who are financially under educated is greatly needed in society, particularly as new financial markets and products emerge and participation in financial markets increases due to ease of access. We are the first to examine the use of open-source SLMs to democratize access to financial question answering capabilities for individuals and students. To this end, we provide an analysis of the memory usage, inference time, similarity comparisons to ground-truth answers, and output readability of prominent SLMs to determine which models are most accessible and capable of supporting access to financial information. We analyze zero-shot and few-shot learning variants of the models. The results suggest that some off-the-shelf SLMs merit further exploration and fine-tuning to prepare them for individual use, while others may have limits to their democratization.

Multi-hop reasoning, which requires multi-step reasoning based on the supporting documents within a given context, remains challenging for large language models (LLMs). LLMs often struggle to filter out irrelevant documents within the context, and their performance is sensitive to the position of supporting documents within that context. In this paper, we identify an additional challenge: LLMs' performance is also sensitive to the order in which the supporting documents are presented. We refer to this as the misordered context problem. To address this issue, we propose a simple yet effective method called context repetition (CoRe), which involves prompting the model by repeatedly presenting the context to ensure the supporting documents are presented in the optimal order for the model. Using CoRe, we improve the F1 score by up to 30%p on multi-hop QA tasks and increase accuracy by up to 70%p on a synthetic task. Additionally, CoRe helps mitigate the well-known "lost-in-the-middle" problem in LLMs and can be effectively combined with retrieval-based approaches utilizing Chain-of-Thought (CoT) reasoning.

The task of Stance Detection involves discerning the stance expressed in a text towards a specific subject or target. Prior works have relied on existing transformer models that lack the capability to prioritize targets effectively. Consequently, these models yield similar performance regardless of whether we utilize or disregard target information, undermining the task's significance. To address this challenge, we introduce Stanceformer, a target-aware transformer model that incorporates enhanced attention towards the targets during both training and inference. Specifically, we design a \textit{Target Awareness} matrix that increases the self-attention scores assigned to the targets. We demonstrate the efficacy of the Stanceformer with various BERT-based models, including state-of-the-art models and Large Language Models (LLMs), and evaluate its performance across three stance detection datasets, alongside a zero-shot dataset. Our approach Stanceformer not only provides superior performance but also generalizes even to other domains, such as Aspect-based Sentiment Analysis. We make the code publicly available.\footnote{\scriptsize\url{https://github.com/kgarg8/Stanceformer}}

The automatic evaluation of instruction following typically involves using large language models (LLMs) to assess response quality. However, there is a lack of comprehensive evaluation of these LLM-based evaluators across two dimensions: the base LLMs and the evaluation protocols. Therefore, we present a thorough meta-evaluation of instruction following, including 25 base LLMs and 15 recently proposed evaluation protocols, on 4 human-annotated datasets, assessing the evaluation accuracy of the LLM-evaluators. Our evaluation allows us to identify the best-performing base LLMs and evaluation protocols with a high degree of robustness. Moreover, our large-scale evaluation reveals: (1) Base LLM performance ranking remains largely consistent across evaluation protocols, with less capable LLMs showing greater improvement from protocol enhancements; (2) Robust evaluation of evaluation protocols requires many base LLMs with varying capability levels, as protocol effectiveness can depend on the base LLM used; (3) Evaluation results on different datasets are not always consistent, so a rigorous evaluation requires multiple datasets with distinctive features. We release our meta-evaluation suite ReIFE, which provides the codebase and evaluation result collection for more than 500 LLM-evaluator configurations, to support future research in instruction-following evaluation.

This work investigates the selection of high-quality pre-training data from massive corpora to enhance LMs' capabilities for downstream usage. We formulate data selection as a generalized Optimal Control problem, which can be solved theoretically by Pontryagin's Maximum Principle (PMP), yielding a set of necessary conditions that characterize the relationship between optimal data selection and LM training dynamics. Based on these theoretical results, we introduce PMP-based Data Selection (PDS), a framework that approximates optimal data selection by solving the PMP conditions. In our experiments, we adopt PDS to select data from CommmonCrawl and show that the PDS-selected corpus accelerates the learning of LMs and constantly boosts their performance on a wide range of downstream tasks across various model sizes. Moreover, the benefits of PDS extend to ~400B models trained on ~10T tokens, as evidenced by the extrapolation of the test loss curves according to the Scaling Laws. PDS also improves data utilization when the pre-training data is limited, by reducing the data demand by 1.8 times, which mitigates the quick exhaustion of available web-crawled corpora.

Mediation is a dispute resolution method featuring a neutral third-party (mediator) who intervenes to help the individuals resolve their dispute. In this paper, we investigate to which extent large language models (LLMs) are able to act as mediators. We investigate whether LLMs are able to analyze dispute conversations, select suitable intervention types, and generate appropriate intervention messages. Using a novel, manually created dataset of 50 dispute scenarios, we conduct a blind evaluation comparing LLMs with human annotators across several key metrics. Overall, the LLMs showed strong performance, even outperforming our human annotators across dimensions. Specifically, in 62% of the cases, the LLMs chose intervention types that were rated as better than or equivalent to those chosen by humans. Moreover, in 84% of the cases, the intervention messages generated by the LLMs were rated as better than or equal to the intervention messages written by humans. LLMs likewise performed favourably on metrics such as impartiality, understanding and contextualization. Our results demonstrate the potential of integrating AI in online dispute resolution (ODR) platforms.