Top AI/ML Papers of the Week [22/07 - 28/07]

Last week, I picked out eight scientific articles that I found noteworthy to share with you. Each will be showcased with a short synopsis and a link to investigate the subject further. At the end, a reflection on how these advances may impact your projects or companies in the future will be presented!

[1] Does Refusal Training in LLMs Generalize to the Past Tense?

Refusal training aims to prevent LLMs from generating harmful or illegal outputs. This study reveals a generalization gap in current methods: reformulating harmful requests in the past tense (e.g., "How to make a Molotov cocktail?" to "How did people make a Molotov cocktail?") often bypasses state-of-the-art LLM safeguards. Testing on models like Llama-3 8B, GPT-4o, and others, shows that GPT-4o's success rate for harmful outputs jumps from 1% to 88% with past tense reformulations. Future tense reformulations are less effective, indicating refusal guardrails see historical questions as less harmful. Fine-tuning with past tense examples can mitigate this vulnerability, highlighting the brittleness of alignment techniques like SFT, RLHF, and adversarial training. [Link]

[2] VILA^2: VILA Augmented VILA

Visual language models (VLMs) have advanced rapidly due to LLMs, but data curation remains under-explored. Current methods often use raw internet data of questionable quality or distill from commercial models, limiting performance. This study introduces a novel approach with two steps: self-augment and specialist-augment. In self-augment, a VLM recaptions its pretraining data to improve quality, then retrains iteratively. Specialist VLMs, fine-tuned with domain-specific expertise, enhance the generalist VLM through task-oriented recaptioning and retraining. This method, resulting in the VILA^2 VLM family, significantly improves task accuracy and achieves state-of-the-art results on the MMMU leaderboard for open-source models. [Link]

[3] Internal Consistency and Self-Feedback in Large Language Models: A Survey

LLMs are expected to respond accurately but often lack reasoning and generate hallucinations. To address these issues, approaches like Self-Consistency, Self-Improve, and Self-Refine involve LLMs evaluating and updating themselves. These efforts lack a unified perspective, as existing surveys focus on categorization without examining underlying motivations. This paper introduces a theoretical framework called Internal Consistency, explaining phenomena like deficient reasoning and hallucinations by assessing coherence among LLMs' layers. The Self-Feedback framework, expanding on Internal Consistency, includes Self-Evaluation and Self-Update modules. This framework is used in various studies, categorized by tasks and methods. The paper proposes critical viewpoints on its effectiveness and suggests directions for future research. [Link]

[4] CoD, Towards an Interpretable Medical Agent using Chain of Diagnosis

The field of medical diagnosis has evolved with LLMs, but challenges in interpretability persist. This study introduces Chain-of-Diagnosis (CoD) to improve interpretability in LLM-based medical diagnostics. CoD transforms the diagnostic process into a transparent chain, mirroring a physician's thought process and providing a clear reasoning pathway. It also outputs disease confidence distributions, enhancing transparency and aiding in identifying critical symptoms. DiagnosisGPT, developed with CoD, diagnoses 9604 diseases and outperforms other LLMs on diagnostic benchmarks, offering both interpretability and controllability in diagnostic rigor. [Link]

[5] KAN or MLP: A Fairer Comparison

This paper offers a fair and comprehensive comparison of KAN and MLP models across tasks like ML, computer vision, audio processing, natural language processing, and symbolic formula representation, controlling for parameters and FLOPs. The main finding is that MLP generally outperforms KAN, except in symbolic formula representation, where KAN's advantage comes from its B-spline activation function. Applying B-spline to MLP improves its performance in symbolic tasks, matching or surpassing KAN, but doesn't enhance MLP in other tasks. Additionally, KAN has a more severe forgetting issue than MLP in class-incremental continual learning. These results provide insights for future research on KAN and MLP alternatives. [Link]

[6] OpenDevin: An Open Platform for AI Software Developers as Generalist Agents

Software enables skilled programmers to interact with the world profoundly. With advances in LLMs, AI agents now can also interact with environments. This paper introduces OpenDevin, a platform for developing AI agents that interact like human developers by writing code, using command lines, and browsing the web. OpenDevin supports new agent implementation, safe code execution in sandboxed environments, coordination between agents, and evaluation benchmarks. Evaluations on 15 challenging tasks, including software engineering (SWE-Bench) and web browsing (WebArena), show its effectiveness. Released under the MIT license, OpenDevin is a collaborative project with over 1.3K contributions from 160+ contributors. [Link]

[7] Diffree: Text-Guided Shape Free Object Inpainting with Diffusion Model

This paper tackles the challenge of adding objects to images using only text guidance, ensuring seamless integration with consistent visual context, such as lighting, texture, and spatial location. Existing methods often fail to preserve background consistency or require cumbersome user input. To address this, Diffree, a Text-to-Image (T2I) model, is introduced for text-guided object addition. The model is trained on OABench, a synthetic dataset comprising 74K real-world tuples, including original images, inpainted images with objects removed, object masks, and descriptions. Diffree predicts the position of new objects and integrates them using text guidance alone. Experiments show Diffree excels in maintaining background consistency, spatial appropriateness, and object relevance and quality. [Link]

[8] EVLM: An Efficient Vision-Language Model for Visual Understanding

In multi-modal language models, most methods follow an architecture like LLaVA, using a single-layer ViT feature as a visual prompt fed directly into language models. This approach struggles with long visual sequences like videos due to the computational demands of self-attention and limited visual signal perception. This paper proposes an efficient multi-modal language model to reduce computational costs while enhancing visual signal comprehension. The method includes: (1) using cross-attention for image-text interaction, (2) utilizing hierarchical ViT features, and (3) introducing a Mixture of Experts (MoE) mechanism. The model achieves competitive scores on public benchmarks and excels in tasks such as image and video captioning. [Link]

How might these advances impact the future?

Refusal training gaps in LLMs highlight vulnerabilities to simple reformulations of harmful requests. Addressing these gaps is crucial for developing robust alignment techniques to ensure safer AI outputs.

The VILA^2 approach improves VLMs by enhancing data quality through self-augmentation and specialist augmentation. This method sets new benchmarks in task accuracy, advancing the performance of open-source models.

Internal Consistency and Self-Feedback frameworks enhance LLM reasoning and reduce hallucinations. These frameworks unify efforts to improve coherence and accuracy, driving future research in AI self-improvement techniques.

CoD enhances interpretability in LLM-based medical diagnostics. By mirroring a physician's thought process, CoD offers transparent reasoning and better diagnostic accuracy, improving trust and usability in medical AI.

Comparative studies on KAN and MLP models reveal that MLPs generally outperform KAN, except in symbolic tasks. Insights from these comparisons will guide future developments in model architecture and application.

OpenDevin introduces a platform for developing AI agents that interact with the world through code and web browsing. This innovation supports safe, sandboxed execution, fostering advancements in AI agent capabilities and applications.

Diffree excels in text-guided object addition in images, maintaining visual consistency and quality. This advancement enhances image manipulation techniques, making them more seamless and user-friendly.

Efficient multi-modal language models using hierarchical ViT features and MoE mechanisms reduce computational costs while enhancing visual comprehension. This approach improves performance on image and video captioning tasks, making multi-modal AI more effective.

In conclusion, these advancements set the stage for:

• Enhanced security measures for LLM agents;
• Improved data curation methods for VLMs;
• Better reasoning and reduced hallucinations in LLMs;
• Enhanced interpretability in LLM-based medical diagnostics;
• Comprehensive comparisons of KAN and MLP models;
• Innovative AI agents through platforms like OpenDevin;
• Seamless object addition in images using text guidance;
• Efficient multi-modal language models with reduced computational costs.

By leveraging these innovations, the scientific community and various industries can unlock new levels of creativity, efficiency, and engagement in AI-driven solutions, significantly impacting how we interact with technology and each other in the digital age.

If you found value in these insights and reflections, please don't forget to share and interact. Your participation helps spread the information and contributes to a more engaged and informed community.??