Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification

Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision understanding, reasoning, and interaction. However, the inference computation and memory increase progressively with the generation of output tokens during decoding, directly affecting the efficacy of MLLMs. Existing methods attempt to reduce the vision context redundancy to achieve efficient MLLMs. Unfortunately, the efficiency benefits of the vision context reduction in the prefill stage gradually diminish during the decoding stage. To address this problem, we proposed a dynamic vision-language context sparsification framework Dynamic-LLaVA, which dynamically reduces the redundancy of vision context in the prefill stage and decreases the memory and computation overhead of the generated language context during decoding. Dynamic-LLaVA designs a tailored sparsification inference scheme for different inference modes, i.e., prefill, decoding with and without KV cache, to achieve efficient inference of MLLMs. In practice, Dynamic-LLaVA can reduce computation consumption by $\sim$75\% in the prefill stage. Meanwhile, throughout the entire generation process of MLLMs, Dynamic-LLaVA reduces the $\sim$50\% computation consumption under decoding without KV cache, while saving $\sim$50\% GPU memory overhead when decoding with KV cache, due to the vision-language context sparsification. Extensive experiments also demonstrate that Dynamic-LLaVA achieves efficient inference for MLLMs with negligible understanding and generation ability degradation or even performance gains compared to the full-context inference baselines. Code is available at https://github.com/Osilly/dynamic_llava .

MoDification: Mixture of Depths Made Easy

Long-context efficiency has recently become a trending topic in serving large language models (LLMs). And mixture of depths (MoD) is proposed as a perfect fit to bring down both latency and memory. In this paper, however, we discover that MoD can barely transform existing LLMs without costly training over an extensive number of tokens. To enable the transformations from any LLMs to MoD ones, we showcase top-k operator in MoD should be promoted to threshold-p operator, and refinement to architecture and data should also be crafted along. All these designs form our method termed MoDification. Through a comprehensive set of experiments covering model scales from 3B to 70B, we exhibit MoDification strikes an excellent balance between efficiency and effectiveness. MoDification can achieve up to ~1.2x speedup in latency and ~1.8x reduction in memory compared to original LLMs especially in long-context applications.

DetectBench: Can Large Language Model Detect and Piece Together Implicit Evidence?

Detecting evidence within the context is a key step in the process of reasoning task. Evaluating and enhancing the capabilities of LLMs in evidence detection will strengthen context-based reasoning performance. This paper proposes a benchmark called DetectBench for verifying the ability to detect and piece together implicit evidence within a long context. DetectBench contains 3,928 multiple-choice questions, with an average of 994 tokens per question. Each question contains an average of 4.55 pieces of implicit evidence, and solving the problem typically requires 7.62 logical jumps to find the correct answer. To enhance the performance of LLMs in evidence detection, this paper proposes Detective Reasoning Prompt and Finetune. Experiments demonstrate that the existing LLMs' abilities to detect evidence in long contexts are far inferior to humans. However, the Detective Reasoning Prompt effectively enhances the capability of powerful LLMs in evidence detection, while the Finetuning method shows significant effects in enhancing the performance of weaker LLMs. Moreover, when the abilities of LLMs in evidence detection are improved, their final reasoning performance is also enhanced accordingly.

Agent Group Chat: An Interactive Group Chat Simulacra For Better Eliciting Collective Emergent Behavior

To investigate the role of language in human collective behaviors, we developed the Agent Group Chat simulation to simulate linguistic interactions among multi-agent in different settings. Agents are asked to free chat in this simulation for their own purposes based on their character setting, aiming to see agents exhibit emergent behaviours that are both unforeseen and significant. Four narrative scenarios, Inheritance Disputes, Law Court Debates, Philosophical Discourses, Movie Casting Contention, are integrated into Agent Group Chat to evaluate its support for diverse storylines. By configuring specific environmental settings within Agent Group Chat, we are able to assess whether agents exhibit behaviors that align with human expectations. We evaluate the disorder within the environment by computing the n-gram Shannon entropy of all the content speak by characters. Our findings reveal that under the premise of agents possessing substantial alignment with human expectations, facilitating more extensive information exchange within the simulation ensures greater orderliness amidst diversity, which leads to the emergence of more unexpected and meaningful emergent behaviors. The code is open source in https://github.com/MikeGu721/AgentGroup, and online platform will be open soon.

Towards the Law of Capacity Gap in Distilling Language Models

Language model (LM) distillation is a trending area that aims to distil the knowledge resided in a large teacher LM to a small student one. While various methods have been proposed to push the distillation to its limits, it is still a pain distilling LMs when a large capacity gap is exhibited between the teacher and the student LMs. The pain is mainly resulted by the curse of capacity gap, which describes that a larger teacher LM cannot always lead to a better student LM than one distilled from a smaller teacher LM due to the affect of capacity gap increment. That is, there is likely an optimal point yielding the best student LM along the scaling course of the teacher LM. Even worse, the curse of capacity gap can be only partly yet not fully lifted as indicated in previous studies. However, the tale is not ever one-sided. Although a larger teacher LM has better performance than a smaller teacher LM, it is much more resource-demanding especially in the context of recent large LMs (LLMs). Consequently, instead of sticking to lifting the curse, leaving the curse as is should be arguably fine. Even better, in this paper, we reveal that the optimal capacity gap is almost consistent across different student scales and architectures, fortunately turning the curse into the law of capacity gap. The law later guides us to distil a 3B student LM (termed MiniMA) from a 7B teacher LM (adapted LLaMA2-7B). MiniMA is demonstrated to yield a new compute-performance pareto frontier among existing 3B LMs on commonly used benchmarks, and its instruction-tuned version (termed MiniChat) outperforms a wide range of 3B competitors in GPT4 evaluation and could even compete with several 7B chat models.

Action based Network for Conversation Question Reformulation

Conversation question answering requires the ability to interpret a question correctly. Current models, however, are still unsatisfactory due to the difficulty of understanding the co-references and ellipsis in daily conversation. Even though generative approaches achieved remarkable progress, they are still trapped by semantic incompleteness. This paper presents an action-based approach to recover the complete expression of the question. Specifically, we first locate the positions of co-reference or ellipsis in the question while assigning the corresponding action to each candidate span. We then look for matching phrases related to the candidate clues in the conversation context. Finally, according to the predicted action, we decide whether to replace the co-reference or supplement the ellipsis with the matched information. We demonstrate the effectiveness of our method on both English and Chinese utterance rewrite tasks, improving the state-of-the-art EM (exact match) by 3.9\% and ROUGE-L by 1.0\% respectively on the Restoration-200K dataset.

CalQNet -- Detection of Calibration Quality for Life-Long Stereo Camera Setups

Many mobile robotic platforms rely on an accurate knowledge of the extrinsic calibration parameters, especially systems performing visual stereo matching. Although a number of accurate stereo camera calibration methods have been developed, which provide good initial "factory" calibrations, the determined parameters can lose their validity over time as the sensors are exposed to environmental conditions and external effects. Thus, on autonomous platforms on-board diagnostic methods for an early detection of the need to repeat calibration procedures have the potential to prevent critical failures of crucial systems, such as state estimation or obstacle detection. In this work, we present a novel data-driven method to estimate the calibration quality and detect discrepancies between the original calibration and the current system state for stereo camera systems. The framework consists of a novel dataset generation pipeline to train CalQNet, a deep convolutional neural network. CalQNet can estimate the calibration quality using a new metric that approximates the degree of miscalibration in stereo setups. We show the framework's ability to predict from a single stereo frame if a state-of-the-art stereo-visual odometry system will diverge due to a degraded calibration in two real-world experiments.