Dylan Zhang

Pengcheng Jiang, Jiacheng Lin, Zhiyi Shi et al. · stanford

Cutting-edge agentic AI systems are built on foundation models that can be adapted to plan, reason, and interact with external tools to perform increasingly complex and specialized tasks. As these systems grow in capability and scope, adaptation becomes a central mechanism for improving performance, reliability, and generalization. In this paper, we unify the rapidly expanding research landscape into a systematic framework that spans both agent adaptations and tool adaptations. We further decompose these into tool-execution-signaled and agent-output-signaled forms of agent adaptation, as well as agent-agnostic and agent-supervised forms of tool adaptation. We demonstrate that this framework helps clarify the design space of adaptation strategies in agentic AI, makes their trade-offs explicit, and provides practical guidance for selecting or switching among strategies during system design. We then review the representative approaches in each category, analyze their strengths and limitations, and highlight key open challenges and future opportunities. Overall, this paper aims to offer a conceptual foundation and practical roadmap for researchers and practitioners seeking to build more capable, efficient, and reliable agentic AI systems.

Alex Sanchez-Stern, Abhishek Varghese, Zhanna Kaufman et al.

Formal verification is a promising method for producing reliable software, but the difficulty of manually writing verification proofs severely limits its utility in practice. Recent methods have automated some proof synthesis by guiding a search through the proof space using a theorem prover. Unfortunately, the theorem prover provides only the crudest estimate of progress, resulting in effectively undirected search. To address this problem, we create QEDCartographer, an automated proof-synthesis tool that combines supervised and reinforcement learning to more effectively explore the proof space. QEDCartographer incorporates the proofs' branching structure, enabling reward-free search and overcoming the sparse reward problem inherent to formal verification. We evaluate QEDCartographer using the CoqGym benchmark of 68.5K theorems from 124 open-source Coq projects. QEDCartographer fully automatically proves 21.4% of the test-set theorems. Previous search-based proof-synthesis tools Tok, Tac, ASTactic, Passport, and Proverbot9001, which rely only on supervised learning, prove 9.6%, 9.8%, 10.9%, 12.5%, and 19.8%, respectively. Diva, which combines 62 tools, proves 19.2%. Comparing to the most effective prior tool, Proverbot9001, QEDCartographer produces 34% shorter proofs 29% faster, on average over the theorems both tools prove. Together, QEDCartographer and non-learning-based CoqHammer prove 30.3% of the theorems, while CoqHammer alone proves 26.6%. Our work demonstrates that reinforcement learning is a fruitful research direction for improving proof-synthesis tools' search mechanisms.

Dylan Zhang, Xuchao Zhang, Chetan Bansal et al.

Major cloud providers have employed advanced AI-based solutions like large language models to aid humans in identifying the root causes of cloud incidents. Despite the growing prevalence of AI-driven assistants in the root cause analysis process, their effectiveness in assisting on-call engineers is constrained by low accuracy due to the intrinsic difficulty of the task, a propensity for LLM-based approaches to hallucinate, and difficulties in distinguishing these well-disguised hallucinations. To address this challenge, we propose to perform confidence estimation for the predictions to help on-call engineers make decisions on whether to adopt the model prediction. Considering the black-box nature of many LLM-based root cause predictors, fine-tuning or temperature-scaling-based approaches are inapplicable. We therefore design an innovative confidence estimation framework based on prompting retrieval-augmented large language models (LLMs) that demand a minimal amount of information from the root cause predictor. This approach consists of two scoring phases: the LLM-based confidence estimator first evaluates its confidence in making judgments in the face of the current incident that reflects its ``grounded-ness" level in reference data, then rates the root cause prediction based on historical references. An optimization step combines these two scores for a final confidence assignment. We show that our method is able to produce calibrated confidence estimates for predicted root causes, validate the usefulness of retrieved historical data and the prompting strategy as well as the generalizability across different root cause prediction models. Our study takes an important move towards reliably and effectively embedding LLMs into cloud incident management systems.

Junlin Yang, Dylan Zhang, Xiangchen Song et al.

We introduce CausaLab, a scalable environment for evaluating interactive causal discovery by LLM agents. Unlike prior evaluations, CausaLab evaluates both whether an agent can solve a problem using causal evidence and whether its answer is supported by a correct hypothesis about the underlying causal mechanism. Each episode places an agent in a synthetic laboratory: it receives prior measurement records, intervenes on a manipulator crystal, and predicts the resonance frequency of a held-out reactor crystal governed by the same mechanism. The hidden data-generating process is a randomly sampled structural causal model (SCM), so success requires recovering both a causal graph and structural equations rather than recalling prior knowledge. CausaLab also includes a domain-specific language that records the agent's evolving SCM hypothesis, making trajectories inspectable and comparable with ground truth. Experiments show a persistent gap between prediction and mechanism recovery: in the purely observational 6-node setting, GPT-5.2-high reaches 92% task accuracy but only 0.471 all-edge $F_1$. This observation further motivates our exploration of different interaction strategies: Mixed observation--intervention strategies improve structural fidelity: in the mixed 6-node setting, GPT-5.2-high achieves 80% on both task accuracy and all-edge $F_1$. Yet even strong agents struggle to design informative interventions, as pure intervention strategies perform poorly on both task accuracy and all-edge $F_1$. We identify premature stopping as a major weakness of agents, and show that asking the model to verify the consistency between its hypothesis and past data can help mitigate this issue. CausaLab therefore separates predictive success from causal understanding and exposes current LLM agents' limits as experimental causal reasoners.

Qirun Dai, Xiao Liu, Jiawei Zhang et al.

Despite the importance of causal reasoning, training LLMs to reason causally remains underexplored. Existing data efforts mostly focus on benchmarking LLMs on specific aspects of causality, making them less suitable for training generalizable causal reasoners. To address this, we propose UniCo, a data generation framework that both (1) addresses 18 causal query types across Pearl's Causal Ladder and (2) translates natively symbolic examples into code and natural language forms to simulate real-world use cases where causal terms are not explicitly specified. To ensure data quality, UniCo grounds answers with exact causal inference and filters cases with reasoning shortcuts. Upon supervised finetuning with 66.6K UniCo-generated instances, Qwen3-4B, Qwen3-8B and Olmo-3-7B-Instruct achieve an average of 22.9% improvements across all 18 in-distribution query types, and 8.1% over state-of-the-art causal data generation frameworks on 7 established causal benchmarks outside the training distribution. More importantly, in real-world medical understanding, legal decision, and tabular reasoning, UniCo-trained models consistently display more faithful reasoning traces, outperforming the base models by an average of 20.2% in faithfulness metrics. These suggest that causality-centered training not only strengthens causal reasoning, but also equips LLMs with a causal mindset in general reasoning tasks.

Mohamed Sobhi Jabal, Jikai Zhang, Dominic LaBella et al.

The Brain Tumor Reporting and Data System (BT-RADS) standardizes post-treatment MRI response assessment in patients with diffuse gliomas but requires complex integration of imaging trends, medication effects, and radiation timing. This study evaluates an end-to-end multi-agent large language model (LLM) and convolutional neural network (CNN) system for automated BT-RADS classification. A multi-agent LLM system combined with automated CNN-based tumor segmentation was retrospectively evaluated on 509 consecutive post-treatment glioma MRI examinations from a single high-volume center. An extractor agent identified clinical variables (steroid status, bevacizumab status, radiation date) from unstructured clinical notes, while a scorer agent applied BT-RADS decision logic integrating extracted variables with volumetric measurements. Expert reference standard classifications were established by an independent board-certified neuroradiologist. Of 509 examinations, 492 met inclusion criteria. The system achieved 374/492 (76.0%; 95% CI, 72.1%-79.6%) accuracy versus 283/492 (57.5%; 95% CI, 53.1%-61.8%) for initial clinical assessments (+18.5 percentage points; P<.001). Context-dependent categories showed high sensitivity (BT-1b 100%, BT-1a 92.7%, BT-3a 87.5%), while threshold-dependent categories showed moderate sensitivity (BT-3c 74.8%, BT-2 69.2%, BT-4 69.3%, BT-3b 57.1%). For BT-4, positive predictive value was 92.9%. The multi-agent LLM system achieved higher BT-RADS classification agreement with expert reference standard compared to initial clinical scoring, with high accuracy for context-dependent scores and high positive predictive value for BT-4 detection.

Peiran Li, Fangzhou Lin, Shuo Xing et al.

Autonomous scientific discovery is entering a more dangerous regime: once the evaluator is frozen, a sufficiently strong search process can learn to win the exam without learning the mechanism the task was meant to reveal. This is the idea behind our title. To let the abyss stare back is to make evaluation actively push against the candidate through adaptive falsification, rather than passively certify it through static validation. We introduce DASES, a falsification-driven framework in which an Innovator, an Abyss Falsifier, and a Mechanistic Causal Extractor co-evolve executable scientific artifacts and scientifically admissible counterexample environments under a fixed scientific contract. In a controlled loss-discovery problem with a single editable locus, DASES rejects artifacts that static validation would have accepted, identifies the first candidate that survives the admissible falsification frontier, and discovers FNG-CE, a loss that transfers beyond the synthetic discovery environment and consistently outperforms CE and CE+L2 under controlled comparisons across standard benchmarks, including ImageNet.

Xuying Ning, Katherine Tieu, Dongqi Fu et al.

Recent large language models (LLMs) have demonstrated strong capabilities in understanding and generating code, from competitive programming to repository-level software engineering. In emerging agentic systems, code is no longer only a target output. It increasingly serves as an operational substrate for agent reasoning, acting, environment modeling, and execution-based verification. We frame this shift through the lens of agent harnesses and introduce code as agent harness: a unified view that centers code as the basis for agent infrastructure. To systematically study this perspective, we organize the survey around three connected layers. First, we study the harness interface, where code connects agents to reasoning, action, and environment modeling. Second, we examine harness mechanisms: planning, memory, and tool use for long-horizon execution, together with feedback-driven control and optimization that make harness reliable and adaptive. Third, we discuss scaling the harness from single-agent systems to multi-agent settings, where shared code artifacts support multi-agent coordination, review, and verification. Across these layers, we summarize representative methods and practical applications of code as agent harness, spanning coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalization and recommendation, DevOps, and enterprise workflows. We further outline open challenges for harness engineering, including evaluation beyond final task success, verification under incomplete feedback, regression-free harness improvement, consistent shared state across multiple agents, human oversight for safety-critical actions, and extensions to multimodal environments. By centering code as the harness of agentic AI, this survey provides a unified roadmap toward executable, verifiable, and stateful AI agent systems.

Dylan Zhang, Yanshan Lin, Zhengkun Wu et al.

Learning from past experience benefits from two complementary forms of memory: episodic traces -- raw trajectories of what happened -- and consolidated abstractions distilled across many episodes into reusable, schema-like lessons. Recent agentic-memory systems pursue the consolidated form: an LLM rewrites past trajectories into a textual memory bank that it continuously updates with new interactions, promising self-improving agents without parameter updates. Yet we find that such consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory. We trace the regression to the consolidation step rather than the underlying experience: the same trajectories yield qualitatively different memories under different update schedules, and an episodic-only control that simply retains those trajectories remains competitive with the consolidators we test. In a controlled ARC-AGI Stream environment that exposes Retain, Delete, and Consolidate actions, agents preserve raw episodes by default and double the accuracy of their forced-consolidation counterparts; disabling consolidation entirely (episodic management only) matches this auto regime. Practically, robust agent memory should treat raw episodes as first-class evidence and gate consolidation explicitly rather than firing it after every interaction. Looking forward, reliable agentic memory will require LLMs that can consolidate without overwriting the evidence they depend on.

Dylan Zhang, Curt Tigges, Zory Zhang et al.

This paper investigates the ability of transformer-based models to learn structural recursion from examples. Recursion is a universal concept in both natural and formal languages. Structural recursion is central to the programming language and formal mathematics tasks where symbolic tools currently excel beyond neural models, such as inferring semantic relations between datatypes and emulating program behavior. We introduce a general framework that nicely connects the abstract concepts of structural recursion in the programming language domain to concrete sequence modeling problems and learned models' behavior. The framework includes a representation that captures the general \textit{syntax} of structural recursion, coupled with two different frameworks for understanding their \textit{semantics} -- one that is more natural from a programming languages perspective and one that helps bridge that perspective with a mechanistic understanding of the underlying transformer architecture. With our framework as a powerful conceptual tool, we identify different issues under various set-ups. The models trained to emulate recursive computations cannot fully capture the recursion yet instead fit short-cut algorithms and thus cannot solve certain edge cases that are under-represented in the training distribution. In addition, it is difficult for state-of-the-art large language models (LLMs) to mine recursive rules from in-context demonstrations. Meanwhile, these LLMs fail in interesting ways when emulating reduction (step-wise computation) of the recursive function.

Dylan Zhang, Qirun Dai, Hao Peng

High-quality supervised fine-tuning (SFT) data are crucial for eliciting strong capabilities from pretrained large language models (LLMs). Typically, instructions are paired with multiple responses sampled from other LLMs, which are often out of the distribution of the target model to be fine-tuned. This, at scale, can lead to diminishing returns and even hurt the models' performance and robustness. We propose **GRAPE**, a novel SFT framework that accounts for the unique characteristics of the target model. For each instruction, it gathers responses from various LLMs and selects the one with the highest probability measured by the target model, indicating that it aligns most closely with the target model's pretrained distribution; it then proceeds with standard SFT training. We first evaluate GRAPE with a controlled experiment, where we sample various solutions for each question in UltraInteract from multiple models and fine-tune commonly used LMs like LLaMA3.1-8B, Mistral-7B, and Qwen2.5-7B on GRAPE-selected data. GRAPE significantly outperforms strong baselines, including distilling from the strongest model with an absolute gain of up to 13.8%, averaged across benchmarks, and training on 3x more data with a maximum performance improvement of 17.3%. GRAPE's strong performance generalizes to realistic settings. We experiment with the post-training data used for Tulu3 and Olmo-2. GRAPE outperforms strong baselines trained on 4.5 times more data by 6.1% and a state-of-the-art data selection approach by 3% on average performance. Remarkably, using 1/3 of the data and half the number of epochs, GRAPE enables LLaMA3.1-8B to surpass the performance of Tulu3-SFT by 3.5%.

Hanning Zhang, Pengcheng Wang, Shizhe Diao et al.

Large language models (LLMs) have shown promise in performing complex multi-step reasoning, yet they continue to struggle with mathematical reasoning, often making systematic errors. A promising solution is reinforcement learning (RL) guided by reward models, particularly those focusing on process rewards, which score each intermediate step rather than solely evaluating the final outcome. This approach is more effective at guiding policy models towards correct reasoning trajectories. In this work, we propose an entropy-regularized process reward model (ER-PRM) that integrates KL-regularized Markov Decision Processes (MDP) to balance policy optimization with the need to prevent the policy from shifting too far from its initial distribution. We derive a novel reward construction method based on the theoretical results. Our theoretical analysis shows that we could derive the optimal reward model from the initial policy sampling. Our empirical experiments on the MATH and GSM8K benchmarks demonstrate that ER-PRM consistently outperforms existing process reward models, achieving 1% improvement on GSM8K and 2-3% improvement on MATH under best-of-N evaluation, and more than 1% improvement under RLHF. These results highlight the efficacy of entropy-regularization in enhancing LLMs' reasoning capabilities.

Dylan Zhang, Justin Wang, Francois Charton

Instruction tuning -- fine-tuning a large language model (LLM) on pairs of instructions and desired outcomes -- is an approach that enables pre-trained language models to perform real-world tasks and follow human instructions. Its practical success depends on the model learning a broader set of instructions than those it was trained on. Yet the factors that determine model generalization to such \emph{unseen tasks} are not well understood. %To understand the driving factors of generalization, In this paper, we experiment with string rewrites, a symbolic task that serves as a building block for Turing complete Markov algorithms while allowing experimental control of "inputs" and "instructions". We investigate the trade-off between the number of instructions the model is trained on and the number of training samples provided for each instruction and observe that the diversity of the instruction set determines generalization. Generalization emerges once a diverse enough set of tasks is provided, even though very few examples are provided for each task. Instruction diversity also ensures robustness with respect to non-uniform distributions of instructions in the training set.

Qirun Dai, Dylan Zhang, Jiaqi W. Ma et al.

Selecting appropriate training data is crucial for effective instruction fine-tuning of large language models (LLMs), which aims to (1) elicit strong capabilities, and (2) achieve balanced performance across a diverse range of tasks. Influence-based methods show promise in achieving (1) by estimating the contribution of each training example to the model's predictions, but often struggle with (2). Our systematic investigation reveals that this underperformance can be attributed to an inherent bias where certain tasks intrinsically have greater influence than others. As a result, data selection is often biased towards these tasks, not only hurting the model's performance on others but also, counterintuitively, harms performance on these high-influence tasks themselves. As a remedy, we propose BIDS, a Balanced and Influential Data Selection algorithm. BIDS first normalizes influence scores of the training data, and then iteratively balances data selection by choosing the training example with the highest influence on the most underrepresented task. Experiments with both Llama-3 and Mistral-v0.3 on seven benchmarks spanning five diverse capabilities show that BIDS consistently outperforms both state-of-the-art influence-based algorithms and other non-influence-based selection frameworks. Surprisingly, training on a 15% subset selected by BIDS can even outperform full-dataset training with a much more balanced performance. Our analysis further highlights the importance of both instance-level normalization and iterative optimization of selected data for balanced learning of diverse capabilities.

Dylan Zhang, Justin Wang, Tianran Sun

Existing LMs struggle with proof-oriented programming due to data scarcity, which manifest in two key ways: (1) a lack of sufficient corpora for proof-oriented programming languages such as F*, and (2) the absence of large-scale, project-level proof-oriented implementations that can teach the model the intricate reasoning process when performing proof-oriented programming. We present the first on synthetic data augmentation for project level proof oriented programming for both generation and repair. Our method addresses data scarcity by synthesizing basic proof-oriented programming problems for proficiency in that language; incorporating diverse coding data for reasoning capability elicitation and creating new proofs and repair data within existing repositories. This approach enables language models to both synthesize and repair proofs for function- and repository-level code. We show that our fine-tuned 14B parameter model, PoPilot, can exceed the performance of the models that outperforms GPT-4o in project-level proof-oriented programming by 64% relative margin, and can improve GPT-4o's performance by 54% by repairing its outputs over GPT-4o's self-repair.

Dylan Zhang, Yufeng Xu, Haojin Wang et al.

Post-training of reasoning LLMs is a holistic process that typically consists of an offline SFT stage followed by an online reinforcement learning (RL) stage. However, SFT is often optimized in isolation to maximize SFT performance alone. We show that, after identical RL training, models initialized from stronger SFT checkpoints can significantly underperform those initialized from weaker ones. We attribute this to a mismatch typical in current SFT-RL pipelines: the distribution that generates the offline SFT data can differ substantially from the policy optimized during online RL, which learns from its own rollouts. We propose PEAR (Policy Evaluation-inspired Algorithm for Offline Learning Loss Re-weighting), an SFT-stage method that corrects this mismatch and better prepares the model for RL. PEAR uses importance sampling to reweight the SFT loss, with three variants operating at the token, block, and sequence levels. It can be used to augment standard SFT objectives and incurs little additional training overhead once probabilities for the offline data are collected. We conduct controlled experiments on verifiable reasoning games and mathematical reasoning tasks on Qwen 2.5 and 3 and DeepSeek-distilled models. PEAR consistently improves post-RL performance over canonical SFT, with pass at 8 gains up to a 14.6 percent on AIME2025. Our results suggest that PEAR is an effective step toward more holistic LLM post-training by designing and evaluating SFT with downstream RL in mind rather than in isolation.

Rui Pan, Dylan Zhang, Hanning Zhang et al.

Bilevel optimization has shown its utility across various machine learning settings, yet most algorithms in practice require second-order information, making it challenging to scale them up. Only recently, a paradigm of first-order algorithms has emerged in the theoretical literature, capable of effectively addressing bilevel optimization problems. Nevertheless, the practical efficiency of this paradigm remains unverified, particularly in the context of large language models (LLMs). This paper introduces the first scalable instantiation of this paradigm called ScaleBiO, focusing on bilevel optimization for large-scale LLM data reweighting. By combining with a recently proposed memory-efficient training technique called LISA, our novel algorithm allows the paradigm to scale to $\sim$30B-sized LLMs on $8\times$H100 GPUs, marking the first successful application of bilevel optimization under practical scenarios for large-sized LLMs. Empirically, extensive experiments on data reweighting verify the effectiveness of ScaleBiO for different-scaled models, including Llama-3-8B, Gemma-2-9B, Qwen-2-7B, and Qwen-2.5-32B, where bilevel optimization succeeds in instruction-following and math reasoning tasks, outperforming several popular baselines, including uniform sampling, influence-aware data filtering, and reference-model-based sampling methods. Theoretically, ScaleBiO ensures the optimality of the learned data weights, along with a convergence guarantee matching the conventional first-order bilevel optimization paradigm on smooth and strongly convex objectives.

Dylan Zhang, Shizhe Diao, Xueyan Zou et al.

Preference learning provides a promising solution to address the limitations of supervised fine-tuning (SFT) for code language models, where the model is not explicitly trained to differentiate between correct and incorrect code. Recent findings demonstrate that on-policy data is the key to successful preference learning, where the preference data is collected using the same policy LM being trained. Inspired by this, we propose PLUM, an on-policy $\textbf{P}$reference $\textbf{L}$earning framework A$\textbf{u}$gmented with test cases for code L$\textbf{M}$ s. The framework operates in three key stages: (1) automatic generation of test cases from natural language instructions, (2) creation of a preference data by evaluating candidate code solutions sampled from the policy, which can then be used to (3) train the policy LM. PLUM levitates the need to train reward models, allowing for large scale on-policy and online preference data collation. PLUM is evaluated on both standard benchmarks (HumanEval, MBPP) and more challenging ones (LiveCodeBench), delivering substantial improvements over original SFT'ed models and other execution-feedback-driven approaches. We show PLUM's benefits are consistent across various widely-used code LMs even they have been well-trained with SFT. For example, PLUM increases pass rates by up to 4.8% on average on standard benchmarks and 11.8% on LiveCodeBench, demonstrating its effectiveness and generalizability. We also demonstrate the benefits of on-policy and online preference learning by comprehensive experimentation.