Krafton Presents 20 Papers at ICML 2026, Including Record 10 Main Track Submissions

Track record and venue footprint

ICML 2026 runs July 6 to 11 at COEX in Seoul's Gangnam District. Krafton's 10 main-track submissions cleared double-blind review; the remaining 10 landed on the workshop track. Among South Korean game companies the main-track count is highest; globally, including non-Korean studios, it ranks third.

Cumulative output across ICML, NeurIPS, and ICLR stands at 85 papers. The trajectory: 5 main-track papers at NeurIPS 2023, 8 main-track papers across the three venues in 2024, and 15 papers in 2025 — including three of six ICLR acceptances designated as Spotlight Papers. ICML 2026 produces the steepest single-year increment, with the main-track share doubling versus 2025's combined total.

Research scope

The accepted work maps onto the stack required to build foundation models internally: world models, multimodal large language models, preference learning, reasoning, and optimization. Named contributions include generation-accuracy improvements for diffusion language models; an empirical comparison of RLHF and Direct Preference Optimization for human-preference alignment; mitigation of cognitive biases in multimodal LLM evaluation; identification of token correspondence inside world models; and methods for analyzing and improving LLM reasoning processes.

Concurrent with the conference, Krafton and Odyssey, a generative-video technology company, co-hosted an "AI for Games" social event. Approximately 500 attendees from academia and industry participated; presenters and panelists included researchers from Sony AI, Microsoft Research, Odyssey, NC AI, and Nvidia. Discussion tracks spanned game-playing cooperative and competitive agents, real-time controllable generative world models, and automation of QA and content-creation tooling.

Audit priorities

Three of the named contributions carry claims the press summary does not settle. The RLHF-versus-DPO study lacks stated evaluation protocol: whether the comparison isolates alignment quality from downstream task performance, or holds compute and data budgets fixed, is unconfirmed. Alignment comparisons with mismatched budgets are a recurring source of overstated conclusions. The world-model token-correspondence work similarly omits the substrate — image tokens, video latents, or structured state — and substrate materially changes what correspondence denotes. The LLM-reasoning paper's category is opaque: a mechanistic probe, a behavioral trace, and a fine-tuning intervention are not interchangeable, and the distinction determines whether the work advances interpretability or merely catalogues failure modes.

Practitioners implementing these findings should prioritize the diffusion-LLM accuracy work and the RLHF/DPO comparison. Replication gates for the latter must include matched compute, matched data mix, and shared evaluation surfaces before any preference-learning claim is treated as transferable. Capital-allocation channels for compute-heavy research programs extend well beyond grants; instruments such as a yield-bearing stablecoin testnet remain peripheral to the empirical content of these papers but factor into lab-scale economics.