Machine intelligence

Google is at the forefront of innovation in Machine Intelligence, with active research exploring virtually all aspects of machine learning, including deep learning and more classical algorithms. Exploring theory as well as application, much of our work on language, speech, translation, visual processing, ranking and prediction relies on Machine Intelligence. In all of those tasks and many others, we gather large volumes of direct or indirect evidence of relationships of interest, applying learning algorithms to understand and generalize.

Machine Intelligence at Google raises deep scientific and engineering challenges, allowing us to contribute to the broader academic research community through technical talks and publications in major conferences and journals. Contrary to much of current theory and practice, the statistics of the data we observe shifts rapidly, the features of interest change as well, and the volume of data often requires enormous computation capacity. When learning systems are placed at the core of interactive services in a fast changing and sometimes adversarial environment, combinations of techniques including deep learning and statistical models need to be combined with ideas from control and game theory.

Recent Publications

Preview abstract Modern transformers have been scaled up in size, data, training-time compute, and more recently test-time compute based on Chain-of-Thought (CoT), but the consequences of trading off these parameters are unknown. We construct explicit tradeoffs between two of these aspects: the number of layers in a transformer and the number of CoT tokens. Our first result is general: with mild assumptions and a single-token prompt, for any integer k, a transformer using soft attention and SwiGLU MLP layers with L layers and n CoT tokens can be converted to a transformer with at most ckL layers and ⌈n/k⌉ CoT tokens, and vice versa, although with an increase to the size of other parameters. We then a multi-token prompt path following task, finding that additional preprocessing from the multi-token prompt changes the depth-CoT relationship in this case. We construct an attention-only network with L layers and 2^(L−2) heads that can traverse 2^(L−1) − 1 edges along a path with a single CoT token, demonstrating at least an exponential relationship between increases in depth and reductions in the number of CoT tokens needed to traverse the entire path. This relationship for the path following task holds empirically as well, as demonstrated by training networks with up to 5 layers, with the tradeoff between depth and the maximum number of edge traversals in a single CoT token exactly matching our theoretical constructions. View details
Preview abstract When an AI language model receives two conflicting documents, how does it decide which one to trust? Does it rely on how the sources are framed or the physical order of the documents? We evaluated this behavior on Google's Gemma 4-e4b model using a completely counterbalanced experimental design. This setup allowed us to mathematically isolate the specific effects of source framing and reading position, while ensuring the model's natural vocabulary biases were canceled out. Across ten test conditions, we discovered that source framing heavily overpowers reading position by a roughly 4-to-1 margin. View details
Claim-Gated Source-Risk Auditing for Generative Search
Chuhong Xu
Mark Zhou
Iris Qian
Zhaoyi Li
2026
Preview abstract Generative search can cite a passage accurately while omitting a material relationship that changes how the passage should be interpreted. This paper introduces an original methodological framework and implementation blueprint for auditing that gap; it is not an AI detector, classifier, or empirical performance study. The framework combines a five-dimensional query–source–answer record, an operational source-role taxonomy, a derived omission endpoint, and explicit rules for uncertainty and human review. Its central contribution is a claim-gated evaluation protocol specifying the sampling, evidence capture, blinded annotation, adjudication, agreement, accuracy, calibration, error analysis, and paired intervention tests required before detector, prevalence, or defense-effect claims are admissible. A 1,000-row synthetic workbook serves only as a specification fixture for testing serialization and machine-checkable invariants. Its score vectors are normalized within 5 × 10−4 and its recorded labels match their vector argmax, but it lacks the provenance and independent labels required for semantic validation. We therefore make no claim about Web prevalence, detector accuracy, or deployment benefit. The result is an implementation-oriented audit methodology with falsifiable evidence gates for future realWeb evaluation. View details
Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery
Zach Cutts
John Coller
Rachel Chen
Hua Tang
Ryder Easterlin
Dina Zverinski
Petar Sirkovic
arXiv preprint arXiv:2610.24806 (2026)
Preview abstract Despite breakthroughs in AI-driven molecular generation, existing discovery pipelines operate largely in silico, with wet-lab feedback reserved for validation rather than iterative improvement. Whether large language model (LLM)-based reasoning agents can accelerate de novo discovery by iterating with lab data remains unproven. Here we use protein-binding aptamer design as a sandbox (1-2 week cycles, $1-$3 per candidate) to demonstrate a wet lab in a closed loop with AI systems. We tasked agentic pipelines with developing aptamer binders against α-thrombin and myoglobin across three design-test-improve cycles (∼1,000 candidates per round; ∼4,400 constructs). Statistically significant, bead-corrected on-target hit rates among new candidates rose across rounds from 0.0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3 on α-thrombin (37.1% excluding literature motif descendants), and from 10.2% in Round 1 to 3.7% in Round 2 and 33.5% in Round 3 on myoglobin. Single-clone bead-binding titrations of eight full-length α-thrombin candidates confirmed concordance with pooled sequencing enrichment (Pearson r=0.90). In these assays, literature-derived motifs engineered by the agents to reduce off-target reactivity in the pooled screen reached apparent affinities of KD=24.6--133 nM, whereas de novo scaffolds showed low-micromolar binding (KD>1 μM). We also report how our pipelines used myoglobin binding data to exploit the G-quadruplex binder space, diagnose cross-reactivity as on-target specificity dropped from 75.0% to 19.1%, and recover by developing strategies that restored specificity to 72.4%. We demonstrate that empirical lab data is more valuable than existing computational proxies, showing how iterative experimental feedback enables AI agents to diagnose assay artifacts and refine molecular designs across rounds. View details
Preview abstract Published results for log anomaly detection often depend on how the data are split and which preprocessing artifacts are available at test time. We examine this depen- dence by running six pipelines on the Hadoop Distributed File System (HDFS) and Blue Gene/L (BGL) logs under several split, parser, temporal, and target-access settings. The pipelines include three count classifiers, a compact sequence model, and two classifiers that share a semantic representation. We also measure parsing, representation fitting, classifier training, artifact size, and classifier-only inference as separate components. Random splitting gives near-ceiling scores on both datasets. When HDFS count- vector groups are held out intact, or when BGL is split at a chronological cutoff, the scores fall and the model ordering changes. In cross-system evaluation, giving the semantic representation access to an unlabeled target cat- alog raises XGBoost mean average precision from 0.191 to 0.325 for HDFS-to-BGL and from 0.089 to 0.142 for BGL-to-HDFS. At a five-percent review budget, the corre- sponding mean precision and recall are 0.318 and 0.157 in the first direction and 0.129 and 0.221 in the second. The study quantifies how protocol and cost boundaries shape conclusions drawn from these two benchmarks. Code, split indices, predictions, sensitivity runs, and regenera- tion checks accompany the submission. View details
Preview abstract This paper provides an overview of Google's TPUs across five generations, from TPU v2 to Ironwood, highlighting their evolution as scalable, resilient, and sustainable supercomputers for AI training. It details the TPU’s stable architecture and microarchitecture, which has surprisingly easily accommodated the rapidly changing deep neural network workloads, such as the rise of Transformers. Key advancements over eight years include 10x increase in HBM capacity and bandwidth per node, a 100x increase in peak node performance, and a 3600x increase in supercomputer performance. The paper also discusses the role of optical circuit switches and built-in self test in enhancing resilience, how TPU’s carbon footprint was reduced by improving embodied carbon emissions per floating point operation and a 30x gain in performance per Watt. It concludes by identifying six features that may well characterize the successful AI accelerators of this decade. View details
×