TAIJI: Textual Anchoring for Immunizing Jailbreak Images in Vision Language Models

Vision Language Models (VLMs) have demonstrated impressive inference capabilities, but remain vulnerable to jailbreak attacks that can induce harmful or unethical responses. Existing defence methods are predominantly white-box approaches that require access to model parameters and extensive modifications, making them costly and impractical for many real-world scenarios. Although some black-box defences have been proposed, they often impose input constraints or require multiple queries, limiting their effectiveness in safety-critical tasks such as autonomous driving. To address these challenges, we propose a novel black-box defence framework called \textbf{T}extual \textbf{A}nchoring for \textbf{I}mmunizing \textbf{J}ailbreak \textbf{I}mages (\textbf{TAIJI}). TAIJI leverages key phrase-based textual anchoring to enhance the model's ability to assess and mitigate the harmful content embedded within both visual and textual prompts. Unlike existing methods, TAIJI operates effectively with a single query during inference, while preserving the VLM's performance on benign tasks. Extensive experiments demonstrate that TAIJI significantly enhances the safety and reliability of VLMs, providing a practical and efficient solution for real-world deployment.

PFDial: A Structured Dialogue Instruction Fine-tuning Method Based on UML Flowcharts

Process-driven dialogue systems, which operate under strict predefined process constraints, are essential in customer service and equipment maintenance scenarios. Although Large Language Models (LLMs) have shown remarkable progress in dialogue and reasoning, they still struggle to solve these strictly constrained dialogue tasks. To address this challenge, we construct Process Flow Dialogue (PFDial) dataset, which contains 12,705 high-quality Chinese dialogue instructions derived from 440 flowcharts containing 5,055 process nodes. Based on PlantUML specification, each UML flowchart is converted into atomic dialogue units i.e., structured five-tuples. Experimental results demonstrate that a 7B model trained with merely 800 samples, and a 0.5B model trained on total data both can surpass 90% accuracy. Additionally, the 8B model can surpass GPT-4o up to 43.88% with an average of 11.00%. We further evaluate models' performance on challenging backward transitions in process flows and conduct an in-depth analysis of various dataset formats to reveal their impact on model performance in handling decision and sequential branches. The data is released in https://github.com/KongLongGeFDU/PFDial.

Dedicated Feedback and Edit Models Empower Inference-Time Scaling for Open-Ended General-Domain Tasks

Inference-Time Scaling has been critical to the success of recent models such as OpenAI o1 and DeepSeek R1. However, many techniques used to train models for inference-time scaling require tasks to have answers that can be verified, limiting their application to domains such as math, coding and logical reasoning. We take inspiration from how humans make first attempts, ask for detailed feedback from others and make improvements based on such feedback across a wide spectrum of open-ended endeavors. To this end, we collect data for and train dedicated Feedback and Edit Models that are capable of performing inference-time scaling for open-ended general-domain tasks. In our setup, one model generates an initial response, which are given feedback by a second model, that are then used by a third model to edit the response. We show that performance on Arena Hard, a benchmark strongly predictive of Chatbot Arena Elo can be boosted by scaling the number of initial response drafts, effective feedback and edited responses. When scaled optimally, our setup based on 70B models from the Llama 3 family can reach SoTA performance on Arena Hard at 92.7 as of 5 Mar 2025, surpassing OpenAI o1-preview-2024-09-12 with 90.4 and DeepSeek R1 with 92.3.

Predicting Large Language Model Capabilities on Closed-Book QA Tasks Using Only Information Available Prior to Training

The GPT-4 technical report from OpenAI suggests that model performance on specific tasks can be predicted prior to training, though methodologies remain unspecified. This approach is crucial for optimizing resource allocation and ensuring data alignment with target tasks. To achieve this vision, we focus on predicting performance on Closed-book Question Answering (CBQA) tasks, which are closely tied to pre-training data and knowledge retention. We address three major challenges: 1) mastering the entire pre-training process, especially data construction; 2) evaluating a model's knowledge retention; and 3) predicting task-specific knowledge retention using only information available prior to training. To tackle these challenges, we pre-train three large language models (i.e., 1.6B, 7B, and 13B) using 560k dollars and 520k GPU hours. We analyze the pre-training data with knowledge triples and assess knowledge retention using established methods. Additionally, we introduce the SMI metric, an information-theoretic measure that quantifies the relationship between pre-training data, model size, and task-specific knowledge retention. Our experiments reveal a strong linear correlation ($\text{R}^2 > 0.84$) between the SMI metric and the model's accuracy on CBQA tasks across models of varying sizes (i.e., 1.1B, 1.6B, 7B, and 13B). The dataset, model, and code are available at https://github.com/yuhui1038/SMI.

Position: Towards a Responsible LLM-empowered Multi-Agent Systems

The rise of Agent AI and Large Language Model-powered Multi-Agent Systems (LLM-MAS) has underscored the need for responsible and dependable system operation. Tools like LangChain and Retrieval-Augmented Generation have expanded LLM capabilities, enabling deeper integration into MAS through enhanced knowledge retrieval and reasoning. However, these advancements introduce critical challenges: LLM agents exhibit inherent unpredictability, and uncertainties in their outputs can compound across interactions, threatening system stability. To address these risks, a human-centered design approach with active dynamic moderation is essential. Such an approach enhances traditional passive oversight by facilitating coherent inter-agent communication and effective system governance, allowing MAS to achieve desired outcomes more efficiently.

FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model

Large language models have been widely applied, but can inadvertently encode sensitive or harmful information, raising significant safety concerns. Machine unlearning has emerged to alleviate this concern; however, existing training-time unlearning approaches, relying on coarse-grained loss combinations, have limitations in precisely separating knowledge and balancing removal effectiveness with model utility. In contrast, we propose Fine-grained Activation manipuLation by Contrastive Orthogonal uNalignment (FALCON), a novel representation-guided unlearning approach that leverages information-theoretic guidance for efficient parameter selection, employs contrastive mechanisms to enhance representation separation, and projects conflict gradients onto orthogonal subspaces to resolve conflicts between forgetting and retention objectives. Extensive experiments demonstrate that FALCON achieves superior unlearning effectiveness while maintaining model utility, exhibiting robust resistance against knowledge recovery attempts.

MRWeb: An Exploration of Generating Multi-Page Resource-Aware Web Code from UI Designs

Multi-page websites dominate modern web development. However, existing design-to-code methods rely on simplified assumptions, limiting to single-page, self-contained webpages without external resource connection. To address this gap, we introduce the Multi-Page Resource-Aware Webpage (MRWeb) generation task, which transforms UI designs into multi-page, functional web UIs with internal/external navigation, image loading, and backend routing. We propose a novel resource list data structure to track resources, links, and design components. Our study applies existing methods to the MRWeb problem using a newly curated dataset of 500 websites (300 synthetic, 200 real-world). Specifically, we identify the best metric to evaluate the similarity of the web UI, assess the impact of the resource list on MRWeb generation, analyze MLLM limitations, and evaluate the effectiveness of the MRWeb tool in real-world workflows. The results show that resource lists boost navigation functionality from 0% to 66%-80% while facilitating visual similarity. Our proposed metrics and evaluation framework provide new insights into MLLM performance on MRWeb tasks. We release the MRWeb tool, dataset, and evaluation framework to promote further research.

Diverging Preferences: When do Annotators Disagree and do Models Know?

We examine diverging preferences in human-labeled preference datasets. We develop a taxonomy of disagreement sources spanning 10 categories across four high-level classes -- task underspecification, response style, refusals, and annotation errors. We find that the majority of disagreements are in opposition with standard reward modeling approaches, which are designed with the assumption that annotator disagreement is noise. We then explore how these findings impact two areas of LLM development: reward modeling and evaluation. In our experiments, we demonstrate how standard reward modeling methods, like the Bradley-Terry model, fail to differentiate whether a given preference judgment is the result of unanimous agreement among annotators or the majority opinion among diverging user preferences. We also find that these tendencies are also echoed by popular LLM-as-Judge evaluation methods, which consistently identify a winning response in cases of diverging preferences. These findings highlight remaining challenges in LLM evaluations, which are greatly influenced by divisive features like response style, and in developing pluralistically aligned LLMs. To address these issues, we develop methods for identifying diverging preferences to mitigate their influence on evaluation and training.

HelpSteer2-Preference: Complementing Ratings with Preferences

Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. We also demonstrate the effectiveness of this reward model at aligning models to follow instructions in RLHF. We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2 and openly release the trained Reward Model at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward

Adaptive Guardrails For Large Language Models via Trust Modeling and In-Context Learning

Guardrails have become an integral part of Large language models (LLMs), by moderating harmful or toxic response in order to maintain LLMs' alignment to human expectations. However, the existing guardrail methods do not consider different needs and access rights of individual users, and treat all the users with the same rule. This study introduces an adaptive guardrail mechanism, supported by trust modeling and enhanced with in-context learning, to dynamically modulate access to sensitive content based on user trust metrics. By leveraging a combination of direct interaction trust and authority-verified trust, the system precisely tailors the strictness of content moderation to align with the user's credibility and the specific context of their inquiries. Our empirical evaluations demonstrate that the adaptive guardrail effectively meets diverse user needs, outperforming existing guardrails in practicality while securing sensitive information and precisely managing potentially hazardous content through a context-aware knowledge base. This work is the first to introduce trust-oriented concept within a guardrail system, offering a scalable solution that enriches the discourse on ethical deployment for next-generation LLMs.