Focused Large Language Models are Stable Many-Shot Learners

In-Context Learning (ICL) enables large language models (LLMs) to achieve rapid task adaptation by learning from demonstrations. With the increase in available context length of LLMs, recent experiments have shown that the performance of ICL does not necessarily scale well in many-shot (demonstration) settings. We theoretically and experimentally confirm that the reason lies in more demonstrations dispersing the model attention from the query, hindering its understanding of key content. Inspired by how humans learn from examples, we propose a training-free method FocusICL, which conducts triviality filtering to avoid attention being diverted by unimportant contents at token-level and operates hierarchical attention to further ensure sufficient attention towards current query at demonstration-level. We also design an efficient hyperparameter searching strategy for FocusICL based on model perplexity of demonstrations. Comprehensive experiments validate that FocusICL achieves an average performance improvement of 5.2% over vanilla ICL and scales well with many-shot demonstrations.

Poor-Supervised Evaluation for SuperLLM via Mutual Consistency

The guidance from capability evaluations has greatly propelled the progress of both human society and Artificial Intelligence. However, as LLMs evolve, it becomes challenging to construct evaluation benchmarks for them with accurate labels on hard tasks that approach the boundaries of human capabilities. To credibly conduct evaluation without accurate labels (denoted as poor-supervised evaluation), we propose the PoEM framework. We first prove that the capability of a model can be equivalently assessed by the consistency between it and certain reference model, when their prediction distributions are independent and the sample size is infinite. To alleviate the insufficiencies of the conditions in reality, we further introduce an algorithm that treats humans (when available) and the models under evaluation as reference models, alternately conducting model weights calibration and filtering during E-step and M-step. Comprehensive experiments across 3 types of tasks with 16 mainstream LLMs have shown that PoEM under poor supervision can achieve an average of 0.98 Pearson correlation coefficient with supervised evaluation results, demonstrating good effectiveness, efficiency and generalizability. More generally, PoEM has advanced the evaluation paradigm evolution from human-centric to human&model-centric by treating both of them as reference models, mitigating the limitations of human evaluation in the era of LLMs.

Make Every Penny Count: Difficulty-Adaptive Self-Consistency for Cost-Efficient Reasoning

Self-consistency (SC), a widely used decoding strategy for chain-of-thought reasoning, shows significant gains across various multi-step reasoning tasks but comes with a high cost due to multiple sampling with the preset size. Its variants, Adaptive self-consistency (ASC) and Early-stopping self-consistency (ESC), dynamically adjust the number of samples based on the posterior distribution of a set of pre-samples, reducing the cost of SC with minimal impact on performance. Both methods, however, do not exploit the prior information about question difficulty. It often results in unnecessary repeated sampling for easy questions that could be accurately answered with just one attempt, wasting resources. To tackle this problem, we propose Difficulty-Adaptive Self-Consistency (DSC), which leverages the difficulty information from both prior and posterior perspectives to adaptively allocate inference resources, further reducing the cost of SC. To demonstrate the effectiveness of DSC, we conduct extensive experiments on three popular categories of reasoning tasks: arithmetic, commonsense and symbolic reasoning on six benchmarks. The empirical results show that DSC consistently surpasses the strong baseline ASC and ESC in terms of costs by a significant margin, while attaining comparable performances.

Examining the Commitments and Difficulties Inherent in Multimodal Foundation Models for Street View Imagery

The emergence of Large Language Models (LLMs) and multimodal foundation models (FMs) has generated heightened interest in their applications that integrate vision and language. This paper investigates the capabilities of ChatGPT-4V and Gemini Pro for Street View Imagery, Built Environment, and Interior by evaluating their performance across various tasks. The assessments include street furniture identification, pedestrian and car counts, and road width measurement in Street View Imagery; building function classification, building age analysis, building height analysis, and building structure classification in the Built Environment; and interior room classification, interior design style analysis, interior furniture counts, and interior length measurement in Interior. The results reveal proficiency in length measurement, style analysis, question answering, and basic image understanding, but highlight limitations in detailed recognition and counting tasks. While zero-shot learning shows potential, performance varies depending on the problem domains and image complexities. This study provides new insights into the strengths and weaknesses of multimodal foundation models for practical challenges in Street View Imagery, Built Environment, and Interior. Overall, the findings demonstrate foundational multimodal intelligence, emphasizing the potential of FMs to drive forward interdisciplinary applications at the intersection of computer vision and language.

CogLM: Tracking Cognitive Development of Large Language Models

Piaget's Theory of Cognitive Development (PTC) posits that the development of cognitive levels forms the foundation for human learning across various abilities. As Large Language Models (LLMs) have recently shown remarkable abilities across a wide variety of tasks, we are curious about the cognitive levels of current LLMs: to what extent they have developed and how this development has been achieved. To this end, we construct a benchmark CogLM (Cognitive Ability Evaluation for Language Model) based on PTC to assess the cognitive levels of LLMs. CogLM comprises 1,220 questions spanning 10 cognitive abilities crafted by more than 20 human experts, providing a comprehensive testbed for the cognitive levels of LLMs. Through extensive experiments across multiple mainstream LLMs with CogLM, we find that: (1) Human-like cognitive abilities have emerged in advanced LLMs (GPT-4), comparable to those of a 20-year-old human. (2) The parameter size and optimization objective are two key factors affecting the cognitive levels of LLMs. (3) The performance on downstream tasks is positively correlated with the level of cognitive abilities. These findings fill the gap in research on the cognitive abilities of LLMs, tracing the development of LLMs from a cognitive perspective and guiding the future direction of their evolution.

Biomedical SAM 2: Segment Anything in Biomedical Images and Videos

Medical image segmentation and video object segmentation are essential for diagnosing and analyzing diseases by identifying and measuring biological structures. Recent advances in natural domain have been driven by foundation models like the Segment Anything Model 2 (SAM 2). To explore the performance of SAM 2 in biomedical applications, we designed two evaluation pipelines for single-frame image segmentation and multi-frame video segmentation with varied prompt designs, revealing SAM 2's limitations in medical contexts. Consequently, we developed BioSAM 2, an enhanced foundation model optimized for biomedical data based on SAM 2. Our experiments show that BioSAM 2 not only surpasses the performance of existing state-of-the-art foundation models but also matches or even exceeds specialist models, demonstrating its efficacy and potential in the medical domain.

CityX: Controllable Procedural Content Generation for Unbounded 3D Cities

Generating a realistic, large-scale 3D virtual city remains a complex challenge due to the involvement of numerous 3D assets, various city styles, and strict layout constraints. Existing approaches provide promising attempts at procedural content generation to create large-scale scenes using Blender agents. However, they face crucial issues such as difficulties in scaling up generation capability and achieving fine-grained control at the semantic layout level. To address these problems, we propose a novel multi-modal controllable procedural content generation method, named CityX, which enhances realistic, unbounded 3D city generation guided by multiple layout conditions, including OSM, semantic maps, and satellite images. Specifically, the proposed method contains a general protocol for integrating various PCG plugins and a multi-agent framework for transforming instructions into executable Blender actions. Through this effective framework, CityX shows the potential to build an innovative ecosystem for 3D scene generation by bridging the gap between the quality of generated assets and industrial requirements. Extensive experiments have demonstrated the effectiveness of our method in creating high-quality, diverse, and unbounded cities guided by multi-modal conditions. Our project page: https://cityx-lab.github.io.

Integrate the Essence and Eliminate the Dross: Fine-Grained Self-Consistency for Free-Form Language Generation

Self-consistency (SC), leveraging multiple samples from LLMs, shows significant gains on various reasoning tasks but struggles with free-form generation due to the difficulty of aggregating answers. Its variants, UCS and USC, rely on sample selection or voting mechanisms to improve output quality. These methods, however, face limitations due to their inability to fully utilize the nuanced consensus knowledge present within multiple candidate samples, often resulting in suboptimal outputs. We propose Fine-Grained Self-Consistency (FSC) to addresses these limitations by extracting and integrating segment-level commonalities from candidate samples, enhancing the performance of LLMs both in open-ended and reasoning tasks. Based on this, we present two additional strategies: candidate filtering, which enhances overall quality by identifying highly similar candidate sets, and merging, which reduces input token requirements by combining similar samples. The effectiveness of FSC is demonstrated through extensive experiments on various tasks, including summarization, code generation, and mathematical reasoning, using GPT-3.5-turbo and GPT-4. The results indicate significant improvements over baseline methods, showcasing the potential of FSC to optimize output quality by effectively synthesizing fine-grained consensus knowledge from multiple samples.

Dynamic Stochastic Decoding Strategy for Open-Domain Dialogue Generation

Stochastic sampling strategies such as top-k and top-p have been widely used in dialogue generation task. However, as an open-domain chatting system, there will be two different conversation scenarios, i.e. chit-chat and knowledge-based question answering. In the former situation, responses diversity is essential due to the one-to-many nature in dialogue. The latter, on the other hand, requires less randomness given that stochastic decoding strategy entails the risk of generating incorrect information. As a result, an adaptive and flexible decoding strategy is needed to cope with these two scenarios simultaneously. To this end, we propose the dynamic decoding strategy (DDS), which can adjust the decoding space w.r.t. different contexts. In DDS, both sequence-level and token-level adaptive search can be achieved to adjust the decoding process in a unified framework. Besides, our adaptive algorithm can not only be used during model inference, but it can also be applied during the model training stage to further enhance the performance. Comprehensive experiments indicate that the proposed decoding strategy can consistently improve the performance of pre-trained dialogue models when coupled with four well-used stochastic decoding algorithms.

ALDM-Grasping: Diffusion-aided Zero-Shot Sim-to-Real Transfer for Robot Grasping

To tackle the "reality gap" encountered in Sim-to-Real transfer, this study proposes a diffusion-based framework that minimizes inconsistencies in grasping actions between the simulation settings and realistic environments. The process begins by training an adversarial supervision layout-to-image diffusion model(ALDM). Then, leverage the ALDM approach to enhance the simulation environment, rendering it with photorealistic fidelity, thereby optimizing robotic grasp task training. Experimental results indicate this framework outperforms existing models in both success rates and adaptability to new environments through improvements in the accuracy and reliability of visual grasping actions under a variety of conditions. Specifically, it achieves a 75\% success rate in grasping tasks under plain backgrounds and maintains a 65\% success rate in more complex scenarios. This performance demonstrates this framework excels at generating controlled image content based on text descriptions, identifying object grasp points, and demonstrating zero-shot learning in complex, unseen scenarios.