Can a weaker model judge a stronger one if a critic argues against the answer first? Across five model pairings on verifiable code and logic tasks, debate lifted the judge's reward signal by 7 to 16 points in three pairings and did nothing in two. It helps only when the critic out-classifies the judge and the judge verifies the critique instead of trusting it; a cheap answer, critique, judge loop captures most of the benefit.
Naman Goyal
Researcher at Google DeepMind and part of the founding team of Gemini Deep Research. I work on long-horizon agents and on the alignment and calibration of large language models.
Before DeepMind I worked on pretraining foundation models at NVIDIA, and interned at Apple (multimodal document understanding) and Adobe Research (adversarially robust metric learning). M.S. in Computer Science from Columbia University, Bachelor’s in Computer Science from IIT Ropar, India, where I graduated Institute Rank 1. Most recently an invited speaker and panelist at ICML 2026 in Seoul.
Papers
Reasoning models keep "thinking" after they have already committed to an answer. ProFIL trains one probe on the frozen base model to detect those post-commitment steps, then filters high-theater rollouts inside GRPO. Across four domains and two model families it cuts reasoning theater by 11 to 100 percent and shortens chains without hurting accuracy.
The Gemini 2.5 model family report: thinking models with native multimodality, long context, and the agentic capabilities behind Deep Research. Cited 4,800+ times.
Three on-device experiences built end to end: answering questions about what is on the user's screen (LayoutLMv3 fine-tuned on 100k+ app screenshots with automatically generated labels), auto-filling a form from the previous screen, and smart replies for users who switch languages mid-sentence.
A Survey on Self-Supervised Learning Approaches for Improving Multimodal Representation Learning
Oct 2022
A map of self-supervised methods for multimodal representations: cross-modal generation, cross-modal transformer pre-training, cyclic translation between modalities, and unimodal pseudo-labels.
Images encoded as superpixel graphs beat a naive pixel-to-node graph on CIFAR-10, with node saliency landing where a human would point. Modelling Rubik's cube solving as a graph problem matched, but did not beat, a model-free RL baseline.
BinIM binary-searches the adversarial perturbation instead of fixing a step size. On 1,000 ImageNet images it beats FGSM, BIM, and related gradient attacks on three classifiers, driving the true-label probability to about 2e-9.
Technical writing
Building a self-correcting code factory where two local LLM agents write, test, and debug Python scripts entirely on Apple Silicon. No cloud, no cost, no data leaving your machine.
Bringing Karpathy's autoresearch to Apple Silicon with MLX. Architecture deep dive, real benchmarks, and a guide to running autonomous AI experiments on your Mac.
Transitioning from iPhone to Android while keeping your Mac? Here's how to replace key Apple ecosystem services for seamless cross-platform productivity.
An analysis of Google DeepMind's Gemma 3 technical report, covering architecture, training, evaluation, and limitations.
Selected talks
ICML 2026, Trustworthy AI for Good Workshop
Shift Happens: Robustness and Reliability of Multimodal Foundation Models
ICML 2026, Continual Adaptation at Scale (CATS) Workshop
Shift Happens: Robustness and Reliability of Multimodal Foundation Models
Seoul Forum on AI Safety and Security (SFASS), Frontier AI Red-Teaming Workshop
Robustness of Multimodal Foundation Models with Jenny Ni
Toronto Machine Learning Summit (TMLS) 2026, 10th anniversary
Humans + AI: Collaborative Intelligence for Complex Decision-Making
The AI Conference 2025
The Ascendancy and Challenges of Agentic Large Language Models
Videos
All talks
2026
GDM x GDG AI for Science meetup, Google Korea, Seoul Shift Happens: Robustness and Reliability of Multimodal Foundation Modelsevent
Jul 11
ICML 2026, Pluralistic Alignment Workshop, COEX, Seoul Panel discussion panelGoogle Researchworkshop
Jul 11
ICML 2026, EMM-QA Workshop, COEX, Seoul Multimodal Robustness Under Distribution Shift: Evaluation Frameworks for Real-World AI Deployment with Jenny NiICML pageICML workshopGoogle Researchworkshop
Jul 11
ICML 2026, Trustworthy AI for Good Workshop, COEX Grand Ballroom, Seoul Shift Happens: Robustness and Reliability of Multimodal Foundation ModelsICML pageICML workshopGoogle Researchworkshop
Jul 10
ICML 2026, Continual Adaptation at Scale (CATS) Workshop, COEX, Seoul Shift Happens: Robustness and Reliability of Multimodal Foundation ModelsICML pageICML workshopworkshop
Jul 10
ICML 2026, Muslims in ML (MusIML) Workshop, COEX, Seoul Shift Happens: Robustness and Reliability of Multimodal Foundation ModelsICML talk pageworkshop
Jul 9
ICML 2026, Interactive Demos (Google), COEX, Seoul Shift Happens: Robustness and Reliability of Multimodal Foundation Models with Jenny Ni demoGoogle Research
Jul 9
Seoul Forum on AI Safety and Security (SFASS), Frontier AI Red-Teaming Workshop, Seoul Robustness of Multimodal Foundation Models with Jenny Niforumpress release
Jul 8
Toronto Machine Learning Summit (TMLS) 2026, 10th anniversary, MaRS Discovery District, Toronto Humans + AI: Collaborative Intelligence for Complex Decision-Making workshopWatchsummitannouncement
Jun 19
AGI House, Gemini Build Day, Bay Area Building with Gemini 3.5 Flash and the Antigravity harnessevent
May 30
ML Week, Hybrid AI 2026, The Clift Royal Sonesta, San Francisco Taming Non-Determinism: A Framework for Evaluation and Observability in Autonomous Agent Trajectoriessession
May 5
ODSC AI East 2026, virtual Agentic LLMs in Practice: Tools, Function Calling, and Workflow Design workshopWatchevent
Apr 30
AIAI Agentic AI Summit Silicon Valley 2026, San Jose Taming Non-Determinism: A Framework for Evaluation and Observability in Autonomous Agent TrajectoriesWatchsummitspeaker page
Apr 15
DeveloperWeek 2026, AI DevWorld Expo Stage, San Jose Convention Center The Ascendancy and Challenges of Agentic Large Language Modelssession
Feb 18
2025
The AI Conference 2025, Pier 48, San Francisco The Ascendancy and Challenges of Agentic Large Language ModelsWatchslidesspeaker page
Sep 18
Adobe AEP GenAI Seminar Series, Adobe Research World Headquarters, San Jose From Autonomous Reasoning to Collaborative Ecosystems: Architectures for the Next Generation of Enterprise AI Agents
Sep 16
SecurityWeek AI Risk Summit and CISO Forum, The Ritz-Carlton, Half Moon Bay The Ascendancy and Challenges of Agentic Large Language Modelssession
Aug 20
Jul 29
AI DevSummit 2025, South San Francisco Conference Center The Dual Edge of Multimodal AI: Advancing Accessibility While Navigating Biassession
May 28
Conf42 Machine Learning 2025, virtual Beyond Text Generation: The Rise of Agentic AI and Its Transformative PotentialWatchtalk page
May 8
Judging
Hackathon and competition judging. Open to judging invitations.
May 2026
Google I/O Hackathon, Cerebral Valley x Google DeepMind, SHACK15, San Francisco First-round judge · 100+ builders, $20K+ in prizeseventprojects
May 2026
Apr 2026
Apr 2026
Stanford x DeepMind Hackathon, Build with Google AI Studio, Stanford Judge
Apr 2026
YC x Google DeepMind Multimodal Hackathon, San Francisco Judge · Gemini 3.1, Lyria, and Nano Banana 2 early access; listed with the DeepMind team on the event pageevent
Mar 2026
Press
Exclusive interview during ICML 2026 in Seoul (in Korean).
Press release naming the workshop presenters from Google DeepMind, Google, and Microsoft.
Projects
codex-vision
May 2026
A small Claude Code skill that lets Claude review screenshots, generate UI mocks, and edit images by routing them through OpenAI Codex's vision tools — three modes, one shell command per artifact, gated by a 50-case triggering eval.
MacBook Pro M5 Picker
May 2026
Interactive picker for all 228 MacBook Pro M5 build combinations, with live pricing.
Elsewhere
- ICML 2026 Trustworthy AI for Good workshop, speakers
- ICML 2026 Continual Adaptation at Scale workshop, speakers
- ODSC AI East 2026, speaker profile
- Agentic AI Summit Silicon Valley 2026, speaker page
- The AI Conference 2025, speaker page
- Conf42 Machine Learning 2025, talk page
- TMLS 2026 workshop leads, announcement on LinkedIn