Qian Huang

Michael Moor, Qian Huang, Shirley Wu et al.

Medicine, by its nature, is a multifaceted domain that requires the synthesis of information across various modalities. Medical generative vision-language models (VLMs) make a first step in this direction and promise many exciting clinical applications. However, existing models typically have to be fine-tuned on sizeable down-stream datasets, which poses a significant limitation as in many medical applications data is scarce, necessitating models that are capable of learning from few examples in real-time. Here we propose Med-Flamingo, a multimodal few-shot learner adapted to the medical domain. Based on OpenFlamingo-9B, we continue pre-training on paired and interleaved medical image-text data from publications and textbooks. Med-Flamingo unlocks few-shot generative medical visual question answering (VQA) abilities, which we evaluate on several datasets including a novel challenging open-ended VQA dataset of visual USMLE-style problems. Furthermore, we conduct the first human evaluation for generative medical VQA where physicians review the problems and blinded generations in an interactive app. Med-Flamingo improves performance in generative medical VQA by up to 20\% in clinician's rating and firstly enables multimodal medical few-shot adaptations, such as rationale generation. We release our model, code, and evaluation app under https://github.com/snap-stanford/med-flamingo.

Qian Huang, Jian Vora, Percy Liang et al.

A central aspect of machine learning research is experimentation, the process of designing and running experiments, analyzing the results, and iterating towards some positive outcome (e.g., improving accuracy). Could agents driven by powerful language models perform machine learning experimentation effectively? To answer this question, we introduce MLAgentBench, a suite of 13 tasks ranging from improving model performance on CIFAR-10 to recent research problems like BabyLM. For each task, an agent can perform actions like reading/writing files, executing code, and inspecting outputs. We then construct an agent that can perform ML experimentation based on ReAct framework. We benchmark agents based on Claude v1.0, Claude v2.1, Claude v3 Opus, GPT-4, GPT-4-turbo, Gemini-Pro, and Mixtral and find that a Claude v3 Opus agent is the best in terms of success rate. It can build compelling ML models over many tasks in MLAgentBench with 37.5% average success rate. Our agents also display highly interpretable plans and actions. However, the success rates vary considerably; they span from 100% on well-established older datasets to as low as 0% on recent Kaggle challenges created potentially after the underlying LM was trained. Finally, we identify several key challenges for LM-based agents such as long-term planning and reducing hallucination. Our code is released at https://github.com/snap-stanford/MLAgentBench.

Eric Zelikman, Qian Huang, Gabriel Poesia et al.

Despite recent success in large language model (LLM) reasoning, LLMs struggle with hierarchical multi-step reasoning tasks like generating complex programs. For these tasks, humans often start with a high-level algorithmic design and implement each part gradually. We introduce Parsel, a framework enabling automatic implementation and validation of complex algorithms with code LLMs. With Parsel, we automatically decompose algorithmic tasks into hierarchical natural language function descriptions and then search over combinations of possible function implementations using tests. We show that Parsel can be used across domains requiring hierarchical reasoning, including program synthesis and robotic planning. We find that, using Parsel, LLMs solve more competition-level problems in the APPS dataset, resulting in pass rates over 75\% higher than prior results from directly sampling AlphaCode and Codex, while often using a smaller sample budget. Moreover, with automatically generated tests, we find that Parsel can improve the state-of-the-art pass@1 performance on HumanEval from 67\% to 85\%. We also find that LLM-generated robotic plans using Parsel are more than twice as likely to be considered accurate than directly generated plans. Lastly, we explore how Parsel addresses LLM limitations and discuss how Parsel may be useful for human programmers. We release our code at https://github.com/ezelikman/parsel

Percy Liang, Rishi Bommasani, Tony Lee et al. · stanford

Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios when possible (87.5% of the time). This ensures metrics beyond accuracy don't fall to the wayside, and that trade-offs are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to analyze specific aspects (e.g. reasoning, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, 21 of which were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on the same core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models.

Qian Huang, Hongyu Ren, Jure Leskovec

Few-shot knowledge graph (KG) completion task aims to perform inductive reasoning over the KG: given only a few support triplets of a new relation $\bowtie$ (e.g., (chop,$\bowtie$,kitchen), (read,$\bowtie$,library), the goal is to predict the query triplets of the same unseen relation $\bowtie$, e.g., (sleep,$\bowtie$,?). Current approaches cast the problem in a meta-learning framework, where the model needs to be first jointly trained over many training few-shot tasks, each being defined by its own relation, so that learning/prediction on the target few-shot task can be effective. However, in real-world KGs, curating many training tasks is a challenging ad hoc process. Here we propose Connection Subgraph Reasoner (CSR), which can make predictions for the target few-shot task directly without the need for pre-training on the human curated set of training tasks. The key to CSR is that we explicitly model a shared connection subgraph between support and query triplets, as inspired by the principle of eliminative induction. To adapt to specific KG, we design a corresponding self-supervised pretraining scheme with the objective of reconstructing automatically sampled connection subgraphs. Our pretrained model can then be directly applied to target few-shot tasks on without the need for training few-shot tasks. Extensive experiments on real KGs, including NELL, FB15K-237, and ConceptNet, demonstrate the effectiveness of our framework: we show that even a learning-free implementation of CSR can already perform competitively to existing methods on target few-shot tasks; with pretraining, CSR can achieve significant gains of up to 52% on the more challenging inductive few-shot tasks where the entities are also unseen during (pre)training.

Eric Zelikman, Qian Huang, Percy Liang et al.

Language model training in distributed settings is limited by the communication cost of gradient exchanges. In this short note, we extend recent work from Malladi et al. (2023), using shared randomness to perform distributed fine-tuning with low bandwidth. The method is a natural decentralized extension of memory-efficient Simultaneous Perturbation Stochastic Approximation (SPSA). Each iteration, each machine seeds a Random Number Generator (RNG) to perform local reproducible perturbations on model weights and calculate and exchange scalar projected gradients, which are then used to update each model. By using a (machine, sample) identifier as the random seed, each model can regenerate one another's perturbations. As machines only exchange single-byte projected gradients, this is highly communication efficient. There are also potential privacy benefits, as projected gradients may be calculated on different training data, and models never access the other's data. Our approach not only drastically reduces communication bandwidth requirements but also accommodates dynamic addition or removal of machines during the training process and retains the memory-efficient and inference-only advantages of recent work. We perform proof-of-concept experiments to demonstrate the potential usefulness of this method, building off of rich literature on distributed optimization and memory-efficient training.

Qian Huang, Thijs Willems, King Wang Poon

This study aims to investigate whether GPT-4 can effectively grade assignments for design university students and provide useful feedback. In design education, assignments do not have a single correct answer and often involve solving an open-ended design problem. This subjective nature of design projects often leads to grading problems,as grades can vary between different raters,for instance instructor from engineering background or architecture background. This study employs an iterative research approach in developing a Custom GPT with the aim of achieving more reliable results and testing whether it can provide design students with constructive feedback. The findings include: First,through several rounds of iterations the inter-reliability between GPT and human raters reached a level that is generally accepted by educators. This indicates that by providing accurate prompts to GPT,and continuously iterating to build a Custom GPT, it can be used to effectively grade students' design assignments, serving as a reliable complement to human raters. Second, the intra-reliability of GPT's scoring at different times is between 0.65 and 0.78. This indicates that, with adequate instructions, a Custom GPT gives consistent results which is a precondition for grading students. As consistency and comparability are the two main rules to ensure the reliability of educational assessment, this study has looked at whether a Custom GPT can be developed that adheres to these two rules. We finish the paper by testing whether Custom GPT can provide students with useful feedback and reflecting on how educators can develop and iterate a Custom GPT to serve as a complementary rater.

Qian Huang, King Wang Poon

This study examines how students integrate generative artificial intelligence (AI) into design projects through the lens of the SuperSkillsStack framework, which identifies four key human competencies for effective human-AI collaboration: Agency, Domain Knowledge, Imagination, and Taste. As generative AI increasingly transforms creative practice, design education must consider how human capabilities are cultivated alongside technological tools. Using qualitative thematic analysis, this study analyzes reflective writings from 80 student design teams participating in a human-centered design course. The findings show that students primarily used AI during the early stages of the design process, including brainstorming, information synthesis, and problem framing. However, students consistently relied on human judgment to interpret contextual information, validate AI-generated outputs, and refine design solutions. Domain knowledge derived from field observations enabled students to detect inaccuracies in AI suggestions, while taste played a critical role in evaluating and selecting meaningful ideas. The results suggest that generative AI functions primarily as a cognitive accelerator rather than a replacement for human creativity. The study highlights the importance of cultivating higher-order human capabilities to support effective human-AI collaboration in design education.

Thijs Willems, Sumbul Khan, Qian Huang et al.

This pilot study traces students' reflections on the use of AI in a 13-week foundational design course enrolling over 500 first-year engineering and architecture students at the Singapore University of Technology and Design. The course was an AI-enhanced design course, with several interventions to equip students with AI based design skills. Students were required to reflect on whether the technology was used as a tool (instrumental assistant), a teammate (collaborative partner), or neither (deliberate non-use). By foregrounding this three-way lens, students learned to use AI for innovation rather than just automation and to reflect on agency, ethics, and context rather than on prompt crafting alone. Evidence stems from coursework artefacts: thirteen structured reflection spreadsheets and eight illustrated briefs submitted, combined with notes of teachers and researchers. Qualitative coding of these materials reveals shared practices brought about through the inclusion of Gen-AI, including accelerated prototyping, rapid skill acquisition, iterative prompt refinement, purposeful "switch-offs" during user research, and emergent routines for recognizing hallucinations. Unexpectedly, students not only harnessed Gen-AI for speed but (enabled by the tool-teammate-neither triage) also learned to reject its outputs, invent their own hallucination fire-drills, and divert the reclaimed hours into deeper user research, thereby transforming efficiency into innovation. The implications of the approach we explore shows that: we can transform AI uptake into an assessable design habit; that rewarding selective non-use cultivates hallucination-aware workflows; and, practically, that a coordinated bundle of tool access, reflection, role tagging, and public recognition through competition awards allows AI based innovation in education to scale without compromising accountability.

Qian Huang, Eric Zelikman, Sarah Li Chen et al.

Token embeddings, a mapping from discrete lexical symbols to continuous vectors, are at the heart of any language model (LM). However, lexical symbol meanings can also be determined and even redefined by their structural role in a long context. In this paper, we ask: is it possible for a language model to be performant without \emph{any} fixed token embeddings? Such a language model would have to rely entirely on the co-occurence and repetition of tokens in the context rather than the \textit{a priori} identity of any token. To answer this, we study \textit{lexinvariant}language models that are invariant to lexical symbols and therefore do not need fixed token embeddings in practice. First, we prove that we can construct a lexinvariant LM to converge to the true language model at a uniform rate that is polynomial in terms of the context length, with a constant factor that is sublinear in the vocabulary size. Second, to build a lexinvariant LM, we simply encode tokens using random Gaussian vectors, such that each token maps to the same representation within each sequence but different representations across sequences. Empirically, we demonstrate that it can indeed attain perplexity comparable to that of a standard language model, given a sufficiently long context. We further explore two properties of the lexinvariant language models: First, given text generated from a substitution cipher of English, it implicitly implements Bayesian in-context deciphering and infers the mapping to the underlying real tokens with high accuracy. Second, it has on average 4X better accuracy over synthetic in-context reasoning tasks. Finally, we discuss regularizing standard language models towards lexinvariance and potential practical applications.

Qian Huang, Hongyu Ren, Peng Chen et al.

In-context learning is the ability of a pretrained model to adapt to novel and diverse downstream tasks by conditioning on prompt examples, without optimizing any parameters. While large language models have demonstrated this ability, how in-context learning could be performed over graphs is unexplored. In this paper, we develop \textbf{Pr}etraining \textbf{O}ver \textbf{D}iverse \textbf{I}n-Context \textbf{G}raph S\textbf{y}stems (PRODIGY), the first pretraining framework that enables in-context learning over graphs. The key idea of our framework is to formulate in-context learning over graphs with a novel \emph{prompt graph} representation, which connects prompt examples and queries. We then propose a graph neural network architecture over the prompt graph and a corresponding family of in-context pretraining objectives. With PRODIGY, the pretrained model can directly perform novel downstream classification tasks on unseen graphs via in-context learning. We provide empirical evidence of the effectiveness of our framework by showcasing its strong in-context learning performance on tasks involving citation networks and knowledge graphs. Our approach outperforms the in-context learning accuracy of contrastive pretraining baselines with hard-coded adaptation by 18\% on average across all setups. Moreover, it also outperforms standard finetuning with limited data by 33\% on average with in-context learning.

Xing Li, Qian Huang, Zhijian Wang et al.

Real-time 3D human action recognition has broad industrial applications, such as surveillance, human-computer interaction, and healthcare monitoring. By relying on complex spatio-temporal local encoding, most existing point cloud sequence networks capture spatio-temporal local structures to recognize 3D human actions. To simplify the point cloud sequence modeling task, we propose a lightweight and effective point cloud sequence network referred to as SequentialPointNet for real-time 3D action recognition. Instead of capturing spatio-temporal local structures, SequentialPointNet encodes the temporal evolution of static appearances to recognize human actions. Firstly, we define a novel type of point data, Hyperpoint, to better describe the temporally changing human appearances. A theoretical foundation is provided to clarify the information equivalence property for converting point cloud sequences into Hyperpoint sequences. Secondly, the point cloud sequence modeling task is decomposed into a Hyperpoint embedding task and a Hyperpoint sequence modeling task. Specifically, for Hyperpoint embedding, the static point cloud technology is employed to convert point cloud sequences into Hyperpoint sequences, which introduces inherent frame-level parallelism; for Hyperpoint sequence modeling, a Hyperpoint-Mixer module is designed as the basic building block to learning the spatio-temporal features of human actions. Extensive experiments on three widely-used 3D action recognition datasets demonstrate that the proposed SequentialPointNet achieves competitive classification performance with up to 10X faster than existing approaches.