Yiwen Song

Yiwen Song

Yiwen Song is a research scientist at Google. Her research focuses on Large Language Models (LLMs), particularly at the intersection of multimodality, generative AI, and agentic systems.

Personal Website: https://yiwen-song.github.io/

Authored Publications
Sort By
  • Title
  • Title, descending
  • Year
  • Year, descending
Preview abstract Autonomous research agents can now produce competitive solutions and complete manuscripts, yet their papers routinely contain fabricated citations, method descriptions disconnected from the code, and scores on incorrect scales---failures invisible to evaluations that assess fluency rather than evidentiary grounding. The core problem is verifiability: no existing system maintains a traceable chain from each claim in the paper to its grounding evidence, and current evaluation protocols assess output fluency rather than evidentiary grounding. We address this with Chaine-of-Evidence (CoE), a verifiability standard requiring every claim to trace to its grounding evidence, and instantiate it in Scientist One, an end-to-end research system that maintains evidence chains natively, and CoE Audit, an evaluation protocol with four integrity checks targeting the most damaging chain failures. Auditing 60 papers from four systems, we find every baseline exhibits at least one failure: phantom citations at 4--25%, method-code alignment in at most 2/15 papers. Scientist One achieves zero phantom citations (0/830), the highest alignment rate (7/15), and competitive solver scores. View details
Preview abstract As language models improve and become increasingly deployed to solve a variety of tasks, trustworthiness becomes essential. Calibration is a good proxy for trust: well-calibrated confidence estimates help inform the risk versus reward tradeoff when trusting a specific model output. Unfortunately, even as models improve, they remain poorly calibrated, often biasing towards overconfidence. Additionally, calibration can be gamed: a policy that always predicts the base rate is perfectly calibrated, but completely uninformative. To resolve this, we develop a new metric, expected utility renormalized by the oracle (EURO), that balances calibration and informativeness. We also propose a general-purpose activation-based confidence, utility, and trust estimation protocol (ACUTE) to appropriately adjudicate uncertainty. The ACUTE protocol provides flexible, sample-efficient, and compute-efficient confidence estimators for 3 tasks including multiple choice question answering, tool-calling, and scientific document summarization across 6 models from 4 model families. ACUTE outperforms strong baselines on EURO, while maintaining low calibration error. Taken together, our work shows that equipping LLMs with the ACUTE protocol can improve calibration, utility, and trustworthiness in numerous settings. View details
Preview abstract Time-series forecasting has traditionally been evaluated solely on numerical accuracy, treating models as "black boxes'' that fail to capture the underlying reasoning. To address this gap, we introduce TFRBench, a novel benchmark designed to evaluate the reasoning capabilities of forecasting systems alongside their numerical accuracy. Unlike existing benchmarks, TFRBench requires models to generate verifiable natural language reasoning by analyzing cross-channel dependencies, identifying strategic trends, and justifying significant events using external context. To construct this benchmark, we propose a systematic multi-agent framework comprising Reasoning, Search, Verifier, Forecasting, and Summary agents. Our benchmark spans five diverse domains including Energy, Sales, Web/CloudOps, Transportation, and Finance, covering 10 distinct datasets. Qualitative evaluation confirms that our generated reasoning is highly faithful and effective; specifically, Large Language Models (LLMs) prompted with our generated reasoning demonstrate significantly improved forecasting accuracy compared to direct forecasting with LLMs. Conversely, benchmarking experiments reveal that off-the-shelf LLMs consistently struggle with both reasoning (shows lower LLM-as-Judge scores) and direct numerical forecasting (MAE and MASE), frequently failing to capture domain-specific dynamics. TFRBench thus establishes a new standard for interpretable, reasoning-based evaluation in time-series forecasting. View details
Preview abstract Time-series forecasting has traditionally been evaluated solely on numerical accuracy, treating models as "black boxes'' that fail to capture the underlying reasoning. To address this gap, we introduce TFRBench, a novel benchmark designed to evaluate the reasoning capabilities of forecasting systems alongside their numerical accuracy. Unlike existing benchmarks, TFRBench requires models to generate verifiable natural language reasoning by analyzing cross-channel dependencies, identifying strategic trends, and justifying significant events using external context. To construct this benchmark, we propose a systematic multi-agent framework comprising Reasoning, Search, Verifier, Forecasting, and Summary agents. Our benchmark spans five diverse domains including Energy, Sales, Web/CloudOps, Transportation, and Finance, covering 10 distinct datasets. Qualitative evaluation confirms that our generated reasoning is highly faithful and effective; specifically, Large Language Models (LLMs) prompted with our generated reasoning demonstrate significantly improved forecasting accuracy compared to direct forecasting with LLMs. Conversely, benchmarking experiments reveal that off-the-shelf LLMs consistently struggle with both reasoning (shows lower LLM-as-Judge scores) and direct numerical forecasting (MAE and MASE), frequently failing to capture domain-specific dynamics. TFRBench thus establishes a new standard for interpretable, reasoning-based evaluation in time-series forecasting. View details
Watch and Learn: Learning to use computers from online videos
Luke Song
Yu Su
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2026), pp. 5421-5431
Preview abstract Computer-using agents (CUAs) must plan task workflows across diverse and evolving applications, yet progress is limited by the lack of large-scale, high-quality training data. Existing datasets are narrow, static, and costly to annotate, while synthetic data often yields oversimplified or misaligned behaviors. We present Watch & Learn (W&L), a framework that converts readily available Internet videos of human computer use into executable UI trajectories at scale. Instead of directly generating actions or relying on handcrafted heuristics, we cast trajectory annotation as an inverse dynamics problem that predicts user actions from consecutive screen states, which simplifies learning and generalizes across domains. Through a task-aware retrieval and labeling pipeline, W&L yields over 53K high-quality trajectories that enhance CUAs both as in-context exemplars and as supervised training data. On OSWorld, it consistently improves general-purpose and specialized CUAs, while on WindowsAgentArena it achieves state-of-the-art performance among 7B-scale models under the 15-step limit. These results show that web-scale human demonstration videos can serve as a practical and scalable foundation for advancing real-world CUAs. Project page: https://chanh.ee/wandl/ View details
Preview abstract The exponential growth of machine learning submissions has strained the traditional peer review process, resulting in slow feedback loops for authors and an immense burden on reviewers to rigorously audit technical soundness and verify literature. To address this, we introduce ScholarPeer, a multi-agent framework designed to operationalize the rigorous auditing workflow of a senior researcher. Rather than attempting to replace human judgment, ScholarPeer serves as a co-scientist: acting as a mentor for rapid author iteration prior to submission, and as an active verification assistant that augments human reviewers. The framework structurally decouples contextualization from critique by deploying a sub-domain historian to synthesize the field's trajectory, a baseline scout to proactively hunt for omitted state-of-the-art comparisons, and a multi-aspect Q&A engine that deeply audits technical soundness-scrutinizing internal logical consistency, experimental validity, and mathematical rigor-while cross-referencing claims against top-tier academic venues. We comprehensively evaluate ScholarPeer on ~1,800 ICLR submissions spanning 2020 through 2025. Our results show that ScholarPeer achieves significant win-rates against state-of-the-art fine-tuned models and search-augmented agentic baselines. View details
Preview abstract Synthesizing unstructured research materials into manuscripts is an essential yet under-explored challenge in AI-driven scientific discovery. Existing autonomous writers are rigidly coupled to specific experimental pipelines, and produce superficial literature reviews. We introduce PaperOrchestra, a multi-agent framework for automated AI research paper writing. It flexibly transforms unconstrained pre-writing materials into publication-ready LaTeX manuscripts, including comprehensive literature synthesis and generated visuals, such as plots and conceptual diagrams. To evaluate performance, we present PaperWriting-Eval200 , the first standardized benchmark of reverse-engineered raw materials from 200 top-tier AI conference papers, alongside a suite of automated peer-review evaluators. PaperOrchestra significantly improves literature review depth and citation coverage, outperforming autonomous baselines with 90%–99% win rates and reaching human-level parity. It also outperforms autonomous baselines on overall manuscript quality by 14%–38%, as measured by side-by-side human evaluation. View details
Preview abstract Test-time scaling has shown considerable success in improving the performance of language models on complex reasoning tasks without requiring fine-tuning. However, current strategies, such as self-reflection or ensembling, primarily focus on logical or structural refinement. They do not leverage the guiding potential of affective feedback. Inspired by psychological research showing that emotions can modulate cognitive performance, we introduce HEART--a novel framework that uses emotionally-driven prompts for iterative self-correction. HEART provides feedback on a models' incorrect response using a curated set of concise, emotionally charged phrases based on Paul Ekman's six basic emotions. By systematically varying the emotional tone of the feedback across iterations, our method guides the model to escape flawed reasoning paths and explore more promising alternatives. We evaluate our framework on challenging reasoning benchmarks including OlympiadBench, Humanity's Last Exam, and SimpleQA. Across these benchmarks, our approach delivers significantly deeper reasoning which leads to consistent and significant increase in accuracy compared to existing prompting methods. Crucially, these gains are observed across a diverse range of model architectures, demonstrating the broad applicability of our technique. Overall, our findings suggest that the next frontier in machine reasoning may lie not just in refining logic, but also in understanding and leveraging the 'HEART' of the models. View details
Preview abstract Recently, decomposing complex problems into simple subtasks--a crucial part of human-like natural planning--to solve the given problem has significantly boosted the performance of large language models (LLMs). However, leveraging such planning structures during post-training to boost the performance of smaller open-source LLMs remains underexplored. Motivated by this, we introduce Plan-Tuning, a unified post-training framework that (i) distills synthetic task decompositions (termed “planning trajectories”) from large-scale LLMs and (ii) fine-tunes smaller models via supervised and reinforcement-learning objectives designed to mimic these planning processes to improve complex reasoning. On GSM8k and the MATH benchmarks, plan-tuned models outperform strong baselines by an average ~7%. Furthermore, plan-tuned models show better generalization capabilities on out-of-domain datasets, with average ~10% and ~12% performance improvements on OlympiadBench and AIME 2024, respectively. Our detailed analysis demonstrates how planning trajectories improves complex reasoning capabilities, showing that Plan-Tuning is an effective strategy for improving task-specific performance of smaller LLMs. View details
Preview abstract The proliferation of Large Language Models (LLMs) has opened new opportunities in data science, yet their practical deployment is often constrained by the challenge of discovering relevant data within large and heterogeneous data lakes. Existing approaches, including single-agent and master–slave multi-agent systems, struggle with scalability, information heterogeneity, and robustness to irrelevant files. To address these limitations, we propose a novel multi-agent communication paradigm inspired by the blackboard architecture in traditional AI and software design. In this framework, a central agent posts information requests to a shared blackboard, and autonomous subordinate agents---each responsible for a partition of the data lake---volunteer to respond based on their capabilities. This distributed design improves scalability and flexibility by eliminating the need for a central coordinator to have prior knowledge of agent expertise. We evaluate the approach on three benchmarks that require explicit data discovery: KramaBench and modified versions of DS-Bench and DA-Code to incorporate data discovery. Experimental results demonstrate that the blackboard architecture substantially outperforms baselines, including RAG and the master–slave paradigm, achieving 13% to 57% relative improvement in end-to-end task success and up to a 9% relative gain in F1 score for data discovery across both proprietary and open-source LLMs. These findings establish the blackboard paradigm as a scalable and generalizable communication framework for multi-agent data science systems. View details
×