Vector Institute Brings 73 Papers to ICML 2026, Spanning Reasoning, Video, and Agents

When the International Conference on Machine Learning opened its doors in Seoul on July 6, the delegation from a single Toronto research hub was hard to miss. The Vector Institute arrived at ICML 2026 with 73 accepted papers, 11 of them chosen as spotlights, the largest presence the institute has ever mustered at the field's flagship venue. Spread across the six days of the conference, which ran through July 11, those papers trace a remarkably complete map of where machine learning is heading: reinforcement learning for reasoning, video generation, vision-language systems, autonomous agents, and the long-running project of planning.

It is a haul that says as much about Canada's AI ecosystem as it does about any one result. Vector, founded in 2017 and anchored by faculty across the University of Toronto and partner institutions, has spent a decade building the kind of dense researcher network that produces this volume of accepted work. The 73-paper total is, by the institute's own account, "Vector's strongest showing at ICML to date."

What the delegation brought to Seoul

The spotlight papers cluster around the year's most active problems. In video generation, a team including Vector Faculty Member and NVIDIA VP of AI Research Sanja Fidler presented Motive, short for Motion Attribution for Video Generation, which earned both a spotlight and an Outstanding Paper Honourable Mention. Motive is the first gradient-based framework to attribute the motion in a generated video, as opposed to its appearance, back to specific clips in the training data. That distinction matters for anyone trying to understand or curate what a video model has learned; the authors report that filtering training data down to the highest-attributed clips lifted human preference to a 74.1 percent win rate.

Multimodal reasoning drew a second spotlight. In a paper titled "Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning," Vector Faculty Member Wenhu Chen and collaborators introduced MoCA, a reinforcement learning framework that separates a model's perceptual failures ("bad seeing") from its reasoning failures ("bad thinking") through modality-aware credit assignment. The headline claim is efficiency as much as accuracy: a single 7-billion-parameter model trained this way surpasses GPT-4o on several benchmarks.

The reach extended into the sciences. dnaHNet, a spotlight from Vector Faculty Member Bo Wang's group, is a tokenizer-free, hierarchical foundation model for genomic sequences that uses differentiable dynamic chunking to deliver more than a threefold inference speedup over Transformers while predicting variant fitness zero-shot. Other spotlights took on efficient deployment, with FlexRank turning a single pretrained model into a nested family of shareable sub-models, and domain adaptation for continuous prediction targets.

Agents, reasoning, and a plea for rigor

Two of the fastest-moving areas at ICML this year, post-training for reasoning and autonomous agents, showed up in Vector's non-spotlight papers as well. Work led by Vector Faculty Member Xiaoxiao Li diagnosed a failure mode the authors call the "lazy likelihood-displacement," a collapse that afflicts Group Relative Policy Optimization when it is used to train tool-using search agents; their fix reported gains of up to 45.2 percent over vanilla GRPO. A separate line of work from Freda Shi's group showed that decoupling perception from reasoning during vision-language post-training could raise accuracy while cutting reasoning traces by roughly a fifth.

Not every contribution was a new model. Among the accepted position papers was "Make Planning Research Rigorous Again!", led by Vector Faculty Affiliate Christian Muise, which argues that the surge of interest in getting large language models to plan has ignored six decades of hard-won methodology from the automated planning community. "We believe that this rigor is the cornerstone of thoughtful and reproducible research that can be built upon," the authors write, before laying out benchmarks, tools, and evaluation protocols that they say LLM-planning papers routinely get wrong. Other position papers took up the accountable deployment of agentic systems and fairness in high-stakes decisions, a reminder that the institute is trying to shape not just what gets built but how it is judged.

Taken together, the portfolio is deliberately broad. In its own framing, the institute described work that "spans reinforcement learning and post-training for advanced reasoning, generative AI and video generation, multimodal and vision-language systems, autonomous agents, and planning and foundational work in optimization and machine learning theory," alongside applications reaching into genomics, quantum chemistry, and materials modelling.

Why this matters

A 73-paper count is easy to read as a scoreboard, and at a conference that drew more than 24,000 submissions this year, sheer volume does signal institutional heft. But the more interesting signal is the shape of the portfolio. Vector's spotlights land almost exactly on the frontier the broader field is chasing: attributing and controlling generative video, squeezing frontier-level reasoning out of small multimodal models, and stabilizing the reinforcement learning recipes that agents depend on. When one regional institute independently populates all of those fronts, it is evidence that the research agenda has genuinely converged, and that the talent to pursue it is no longer concentrated in a handful of corporate labs.

It also underscores something about Canada's bet on AI. The Vector model, publicly seeded research capacity tied to universities and industry partners, was designed to keep and grow domestic talent. A record ICML showing is the kind of return that argument was always meant to produce. The presence of position papers on rigor, accountability, and fairness alongside the technical work suggests an institute that wants influence over norms, not just benchmarks.

What to watch

The near-term test is whether the spotlight ideas travel. Motive's motion-level attribution could become standard tooling as video models face mounting questions about their training data; MoCA's perception-versus-reasoning split is a template other multimodal teams will likely borrow. Watch, too, whether Muise's call for planning rigor actually reshapes how LLM-agent papers are reviewed at the next round of conferences. And with ICML's submission numbers still climbing, the question for Vector is no longer whether it can compete on volume, but whether its Toronto network can keep converting scale into the outsized, field-defining results that spotlights are meant to mark.