google-deepmind/concordia v2.1.0

v2.1.0

Repository: google-deepmind/concordia

Tag: v2.1.0

Published: 2025-08-18T21:15:07Z

Prerelease: no

Release notes:

[2.1.0] - 2025-08-18

Changed

• Set randomize choices to false in questionnaires
• Added configurable number of sentences per episode in formative memory generators
• added exponential backoff in retry wrapper
• Increase DEFAULT_MAX_TOKENS.
• Adding fixed acting order to dialogic GM
• Move configurable component preact values above recent observations in

question_of_recent_memories prompt so the former can contextualize the latter e.g. this is the sensible ordering if you pass instructions.

• small improvement to a prompt in the GenerativeClock component
• Modernize the situation_representation_via_narrative component.
• use verbosity and reasoning_effort parameters inside the OpenAI wrapper
• Make formative memories generator throw an error if passed wrong shape parameters

Added

• Add option to remove duplicates, when extracting data from the logs
• Add acting component flag to randomize choices
• Create non-deprecated no_op_context_processor
• An actor and a game master prefabs and required components for running a simulation that follows a strict script. This can be used for generating fine tuning data.
• Parallel stateless questionnaire
• Adding a callback to get the state of the simulation after every step, which can be used to implement custom checkpointing
• Enable loading presaved memory states from agent config
• Marketplace component that handles logic for buyers and sellers trading goods
• questionnaire simulation load memories
• Added a death component
• Add situated_in_time_and_place game master prefab
• Add support for open weights OpenAI models via Together AI.
• Implementing multi-step questionnaire that can handle both open-ended and multiple-choice questions.
• Add option to return raw log from simulation.play

Fixed

• Dummy language model options
• create game_master module in contrib to fix typecheck error
• making OutputTypes explicit strings and adding conversion to and from dictionaries. This enables serialisation.
• fix serialisation to handle action_spec correctly
• Fix action_spec serialization for death gm component
• fix SendEventToRelevantPlayers serialisation to handle action_spec correctly
• Fixing and improving MakeObservation and SendEventToRelevantPlayers by replacing certain llm calls with simple string editing and fixing logic.
• prevent premature termination in default make_observation component
• Minor fix of next acting component, which makes sure the fixed random order starts with the first actor
• OpenAI models no longer support terminators, and remove their hardcoded output limit.
• always use temperature 1.0 for OpenAI since GPT-5 crashes for all other values.
• use max_completion_tokens instead of max_tokens in base_gpt_model.
• Make it so that calling get_currently_active_player on NextActingAllEntities throws a legible error.