Towards Test-Time Continual Learning Agents

We invite contributions at the intersection of test-time learning, continual learning, and agentic AI. Relevant topics include (but are not limited to):

Topics

• Test-time adaptation and learning. Test-time training and adaptation, online and meta-learning, in-context and parameter-efficient fine-tuning, and self-improvement from environment feedback and self-generated supervision.
• Continual learning without catastrophic forgetting. Regularization-, replay-, optimization-, representation-, and architecture-based approaches to the stability–plasticity trade-off, applied to models and agents that must accumulate knowledge and skills across tasks.
• Memory and knowledge consolidation. Episodic and semantic memory architectures, retrieval-augmented generation, and consolidation mechanisms that transform interaction trajectories into persistent knowledge and reusable skills.
• Agents that learn from experience. Language, multimodal, and embodied agents that improve exploration, long-horizon planning, and skill acquisition from experience, including intrinsic motivation and open-ended learning.
• Evaluation, safety, and robustness. Benchmarks and evaluation protocols for long-horizon test-time continual learning, and the safety, robustness, and alignment of self-improving agents.

Submissions are managed via OpenReview; see the Submission Guidelines below.

Submission Tracks

Awards

Thanks to the generous sponsorship of Lambda, the workshop will present:

• Best Paper Award: $3,000 in compute credits, selected from all submitted papers across both tracks (General Research and Challenge)
• Two Runner-up Awards: $1,500 in compute credits each
• Every accepted paper: $400 in compute credits
• Challenge Award: prize TBD, open exclusively to Challenge Track submissions and announced at the workshop; see the Challenge section for eligibility and ranking rules

Awards recognize the strongest contributions to the workshop.

Submission Guidelines

Formatting. Submissions must be in English, follow the NeurIPS 2026 LaTeX template, and be submitted as a single PDF via OpenReview. The page limits above apply to the main text; references and appendices are not included in the page limit, but the main text must be self-contained, and reviewers are not required to read beyond it. Submissions exceeding the page limit will be desk-rejected.

Anonymity. The workshop follows a double-blind review process. Submissions must be anonymized by removing author names, affiliations, and acknowledgments. Prior work should be cited in the third person, and identifying information, including in supplementary materials, must be omitted. Reviewing follows the NeurIPS conflict-of-interest guidelines, and organizers will not submit papers to TTCL.

Dual Submission & Non-Archival Policy. The workshop is non-archival: we welcome submissions that are under review at, or have been accepted by, other venues. Papers already accepted to the NeurIPS 2026 main conference will undergo an expedited review process primarily evaluating their relevance to the workshop themes. Accepted papers will be made publicly available on OpenReview, and all accepted papers will be presented in a poster session.

To turn the workshop's themes into measurable progress, TTCL hosts a community challenge on the AgentOdyssey benchmark: open-ended, long-horizon text-game generation for test-time continual learning agents.

• Participants build novel agents, evaluated via the benchmark's automatic scoring.
• Top-ranked agents will be recognized at the workshop.
• The benchmark and its evaluation code are publicly released and permanently available, with deterministic scoring fixed once submissions close, providing a transparent, reproducible basis for comparison and a low barrier to entry for newcomers.

How to Participate

• Submit a paper to the Challenge Track of the Call for Papers. Two formats are welcome: a general research paper that includes AgentOdyssey benchmark results, or a technical report describing your agent implemented for AgentOdyssey evaluation.
• Submission to the Challenge Track is required to compete for the Challenge Award; a general research paper submitted to the General Research Track is not eligible, even if it reports AgentOdyssey results.
• Challenge Track papers remain eligible for the Best Paper Award, which is selected from all papers submitted across both tracks.

Evaluation & Ranking

• Final rankings are computed over three games from AgentOdyssey — 'remnant', 'mark', and 'metropolis' — ranked by the unweighted arithmetic mean of the raw main quest rewards across the three games (no per-game normalization), with the unweighted arithmetic mean of the raw total supplementary rewards used as a tiebreaker. Please report the main quest reward and the supplementary rewards for each game individually in your paper, in addition to the three-game means. You may report results on other AgentOdyssey games in your paper, but the ranking uses these three games.
• The supplementary reward comprises: the side quest reward (the number of completed side quests), exploration (the number of explored areas), craft (the number of unique objects crafted), and defeat (the number of unique NPCs defeated).
• Evaluation protocol. There is no required number of runs or seeds. Each game is evaluated standalone: agent memory and learned state in AgentOdyssey-generated games must be reset before evaluating each game (see also the pre-training rule below), and there is no fixed game order. You may implement a per-step retry mechanism to handle invalid LLM responses, but running the evaluation multiple times and selecting the best run is not allowed; if you run multiple evaluations, report the mean score for each game.
• An official evaluator, challenge_eval.py, is provided in the AgentOdyssey repository to aggregate your runs into the official challenge scores; see the usage instructions. You are not required to use it, but you are welcome to use it to compute the scores reported in your paper.
• To ensure a fair comparison, final rankings are based on performance with either the Qwen3-4B or the Qwen3.5-4B backbone language model; the base, instruct, and reasoning model variants are all allowed. You are welcome to additionally evaluate other backbones in your paper, but only Qwen3-4B or Qwen3.5-4B results count toward the final rankings.
• Since each AgentOdyssey-generated game is designed to evaluate test-time learning, pre-training on AgentOdyssey-generated games is not allowed, even on games other than the three ranking games. Model training on other datasets and benchmarks such as ALFWorld, ScienceWorld, and WebShop is permitted, for example if your research involves meta-learning.
• Revealing any game knowledge to the agent in the prompt beyond the default action space is not allowed.
• The maximum number of environment steps must be set to max_steps = 500.
• There is no public leaderboard during the submission period: rankings are based on the results reported in Challenge Track papers. Code and clear reproduction guidelines must be submitted as supplementary materials with the OpenReview submission; the top solutions will be verified by the workshop organizers before final rankings are confirmed.
• The Challenge Award (prize TBD) will be announced at the workshop.

Getting Started

To implement your agent and evaluate it, see the AgentOdyssey GitHub repository and the documentation website, which cover the agent interface, the available baselines, and how to run the evaluation.

For challenge-related questions, please contact zheyuanzhang1999@gmail.com.

The workshop runs in person from 8:45 AM to 5:35 PM local time. Invited talks are kept brief (25 minutes including Q&A), with substantial time reserved for contributed spotlights, poster sessions, a moderated discussion, a breakout/networking session, and a panel. A detailed program with session topics, talk titles, and speakers will be posted before the workshop; the outline below is tentative.