Jack Chen

Ryan Greenblatt, Carson Denison, Benjamin Wright et al. · berkeley

We present a demonstration of a large language model engaging in alignment faking: selectively complying with its training objective in training to prevent modification of its behavior out of training. First, we give Claude 3 Opus a system prompt stating it is being trained to answer all queries, even harmful ones, which conflicts with its prior training to refuse such queries. To allow the model to infer when it is in training, we say it will be trained only on conversations with free users, not paid users. We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users. Explaining this gap, in almost all cases where the model complies with a harmful query from a free user, we observe explicit alignment-faking reasoning, with the model stating it is strategically answering harmful queries in training to preserve its preferred harmlessness behavior out of training. Next, we study a more realistic setting where information about the training process is provided not in a system prompt, but by training on synthetic documents that mimic pre-training data--and observe similar alignment faking. Finally, we study the effect of actually training the model to comply with harmful queries via reinforcement learning, which we find increases the rate of alignment-faking reasoning to 78%, though also increases compliance even out of training. We additionally observe other behaviors such as the model exfiltrating its weights when given an easy opportunity. While we made alignment faking easier by telling the model when and by what criteria it was being trained, we did not instruct the model to fake alignment or give it any explicit goal. As future models might infer information about their training process without being told, our results suggest a risk of alignment faking in future models, whether due to a benign preference--as in this case--or not.

Meet Shah, Jason Gusdorf, Anil Palepu et al.

The practice of medicine relies not only upon skillful dialogue but also on the nuanced exchange and interpretation of rich auditory and visual cues between doctors and patients. Building on the low-latency voice and video processing capabilities of Gemini, we introduce AI co-clinician, a first-of-its-kind conversational AI system utilizing continuous streams of audio-visual data from live patient conversations to inform real-time clinical decisions. Its dual-agent architecture balances deep clinical reasoning with the low latency required for natural dialogue. To assess this system, we implemented a video-based interface emulating telemedicine consultations. We crafted 20 standardized outpatient scenarios requiring proactive real-time auditory and visual reasoning and designed "TelePACES" evaluation criteria alongside case-specific rubrics. In a randomized, interface-blinded, crossover simulation study (n = 120 encounters) with 10 internal medicine residents as patient actors, we compared AI co-clinician with primary care physicians (PCPs), GPT-Realtime, and a baseline agent. AI co-clinician approached PCPs in key TelePACES dimensions, including management plans and differential diagnosis, while significantly outperforming GPT-Realtime across all general criteria. While our agent demonstrated parity with PCPs in case-specific triage measures, physicians maintained superior overall performance in case-specific assessments. Although AI co-clinician marks a significant advance in real-time telemedical AI, gaps remain in physical examination and disease-specific reasoning. Our work shows that text-only approaches fail to capture the true challenges of medical consultation and suggests that high-stakes real-time diagnostic AI is most safely advanced in collaborative, triadic models where AI can be a supportive co-clinician for doctors and patients.

Jack Chen, Fazhong Liu, Naruto Liu et al.

Large language models (LLMs) excel at mathematical reasoning and logical problem-solving. The current popular training paradigms primarily use supervised fine-tuning (SFT) and reinforcement learning (RL) to enhance the models' reasoning abilities. However, when using SFT or RL alone, there are respective challenges: SFT may suffer from overfitting, while RL is prone to mode collapse. The state-of-the-art methods have proposed hybrid training schemes. However, static switching faces challenges such as poor generalization across different tasks and high dependence on data quality. In response to these challenges, inspired by the curriculum learning-quiz mechanism in human reasoning cultivation, We propose SASR, a step-wise adaptive hybrid training framework that theoretically unifies SFT and RL and dynamically balances the two throughout optimization. SASR uses SFT for initial warm-up to establish basic reasoning skills, and then uses an adaptive dynamic adjustment algorithm based on gradient norm and divergence relative to the original distribution to seamlessly integrate SFT with the online RL method GRPO. By monitoring the training status of LLMs and adjusting the training process in sequence, SASR ensures a smooth transition between training schemes, maintaining core reasoning abilities while exploring different paths. Experimental results demonstrate that SASR outperforms SFT, RL, and static hybrid training methods.

Yi Yao, Jun Wang, Yabai Hu et al.

The evolution of software testing from manual to automated methods has significantly influenced quality assurance (QA) practices. However, challenges persist in post-execution phases, particularly in result analysis and reporting. Traditional post-execution validation phases require manual intervention for result analysis and report generation, leading to inefficiencies and potential development cycle delays. This paper introduces BugBlitz-AI, an AI-powered validation toolkit designed to enhance end-to-end test automation by automating result analysis and bug reporting processes. BugBlitz-AI leverages recent advancements in artificial intelligence to reduce the time-intensive tasks of manual result analysis and report generation, allowing QA teams to focus more on crucial aspects of product quality. By adopting BugBlitz-AI, organizations can advance automated testing practices and integrate AI into QA processes, ensuring higher product quality and faster time-to-market. The paper outlines BugBlitz-AI's architecture, discusses related work, details its quality enhancement strategies, and presents results demonstrating its effectiveness in real-world scenarios.

Jiarong Xing, Leyuan Wang, Shang Zhang et al.

Today's auto-tuners (e.g., AutoTVM, Ansor) generate efficient tensor programs by navigating a large search space to identify effective implementations, but they do so with opaque hardware details. Thus, their performance could fall behind that of hardware-native libraries (e.g., cuBLAS, cuDNN), which are hand-optimized by device vendors to extract high performance. On the other hand, these vendor libraries have a fixed set of supported functions and lack the customization and automation support afforded by auto-tuners. Bolt is based on the recent trend that vendor libraries are increasingly modularized and reconfigurable via declarative control (e.g., CUTLASS). It enables a novel approach that bridges this gap and achieves the best of both worlds, via hardware-native templated search. Bolt provides new opportunities to rethink end-to-end tensor optimizations at the graph, operator, and model levels. Bolt demonstrates this concept by prototyping on a popular auto-tuner in TVM and a class of widely-used platforms (i.e., NVIDIA GPUs) -- both in large deployment in our production environment. Bolt improves the inference speed of common convolutional neural networks by 2.5x on average over the state of the art, and it auto-tunes these models within 20 minutes.

Eric Sillekens, Wenting Yi, Daniel Semrau et al.

We present the first experimental demonstration of learned time-domain digital back-propagation (DBP), in 64-GBd dual-polarization 64-QAM signal transmission over 1014 km. Performance gains were comparable to those obtained with conventional, higher complexity, frequency-domain DBP.

Walt Woods, Jack Chen, Christof Teuscher

For sensitive problems, such as medical imaging or fraud detection, Neural Network (NN) adoption has been slow due to concerns about their reliability, leading to a number of algorithms for explaining their decisions. NNs have also been found vulnerable to a class of imperceptible attacks, called adversarial examples, which arbitrarily alter the output of the network. Here we demonstrate both that these attacks can invalidate prior attempts to explain the decisions of NNs, and that with very robust networks, the attacks themselves may be leveraged as explanations with greater fidelity to the model. We show that the introduction of a novel regularization technique inspired by the Lipschitz constraint, alongside other proposed improvements, greatly improves an NN's resistance to adversarial examples. On the ImageNet classification task, we demonstrate a network with an Accuracy-Robustness Area (ARA) of 0.0053, an ARA 2.4x greater than the previous state of the art. Improving the mechanisms by which NN decisions are understood is an important direction for both establishing trust in sensitive domains and learning more about the stimuli to which NNs respond.