25 of 25 Reasoning Games Solved, and the Model Never Changed, Team Reports

Researchers at Huawei's Noah's Ark Lab and two British universities report that an agent solved all 25 public ARC-AGI-3 reasoning games without updating its language model's weights. ^{[1] [3]}

The system, Memento 3, completed every level and recorded a mean Relative Human Action Efficiency score of 100.0, according to the team's preprint. ^{[1] [3]} The authors separately report that it used 44% of the human action count. ^{[1] [3]}

The striking part is where the learning happened. Memento 3 kept a revisable rulebook outside the model, converted its rules into executable code and checked that code against what the games actually did. Its underlying language model stayed fixed while the surrounding system changed. ^{[1] [3]}

The paper was submitted to arXiv on October 8, 2026, by researchers affiliated with University College London, Huawei Noah's Ark Lab in the UK and the University of Liverpool. These are author-reported results from a preprint, not an independently replicated finding. The 25 public games are part of a 135-environment benchmark, so the result does not establish performance on unseen tasks or general intelligence. ^{[1] [2] [3]}

The approach uses a language model to build and repair a program that represents how an environment behaves. That gives the agent something it can run to predict outcomes and plan its next move. ^{[1] [3]}

What a perfect score means

ARC-AGI-3 asks players to learn unfamiliar interactive games. They are not given instructions telling them how to play. They must explore, infer the rules and work out a strategy for completing the levels. ^[2]

That makes the task different from answering a question in a chat window. A player has to discover what an action does before deciding whether it helps. Success requires building an understanding of the game from interaction, then using that understanding to reach the goal. ^[2]

ARC Prize, the organisation behind the benchmark, says its human testing involved 458 participants. Each of the 25 public environments was beaten by at least two independent participants. That provides a human reference for both solvability and the number of actions needed to finish. ^[2]

The score measures more than whether an agent eventually gets to the end. According to ARC Prize's scoring description, a 100% result means completing every level at or above the median human action efficiency. An agent therefore needs to succeed without taking too many actions compared with the human baseline. ^[2]

Memento 3's reported 100.0 is that benchmark score, while its action-use figure measures moves relative to humans. The two are not interchangeable. ^{[1] [3]}

Nor is an action count a bill for computing. It measures moves in the environment. Assessing the system's practical cost would also require accounting for the language-model calls and computation used to construct, test and revise its code.

A fixed model can still build a changing system

A language model's weights are the learned parameters that shape its responses. Calling the model frozen means those parameters are not updated during the task. It does not mean every part of the application around it must stay unchanged.

Memento 3 puts its changes in external memory. Its natural-language rulebook records hypotheses about how the environment works, and the system turns those hypotheses into code for predicting the effects of actions and planning what to do. If the predictions fail, it can revise both the rulebook and the program. ^{[1] [3]}

Think of a fixed model working with an instruction manual it can amend. The model does not acquire new weights each time it discovers a rule. Instead, the next decision can draw on a better manual and a better program.

This surrounding software is often called scaffolding. Here it has a specific job: give the language model a way to preserve discoveries, turn them into operational rules and test whether those rules hold. Memento 3's method makes the record of learning explicit rather than leaving it entirely inside a sequence of model responses. ^[3]

The authors say a proposed code update must pass two checks. The language model must judge the code faithful to the written rulebook. The program is also rerun on the moves already played and must produce the same changes in the game that the agent actually observed. In everyday terms, the program must both match the instructions and account for what happened. ^{[1] [3]}

Those checks give an error somewhere useful to go. A failed prediction can trigger a repair to the system's account of the game, rather than merely another attempt at choosing a move. That is the engineering idea worth watching: make experience change an executable model of the task. ^[3]

From discovering rules to running them

Earlier Memento work, as described in the new paper, used external memory to retain experiences and procedures. Memento 3 extends that approach to explicit models of how an environment behaves. It stores an account of what actions will do, which can then support planning. ^[3]

The paper also reports a small Atari Pong case study. A learned feedback controller won 21:0 in each of three evaluated episodes with different openings. The controller is a program that picks each move from what it observes.

While playing, it needed no fresh input from the language model, according to the authors. ^{[1] [3]}

That example shows why converting discoveries into code matters. Once a usable controller exists, the language model need not supply every individual action. The authors report a separation between building the controller and running it. ^[3]

The next test is whether other researchers can reproduce the result and whether the method carries over to held-back environments. Code, run details, the model configuration and computing costs would make that assessment possible. Memento 3 has supplied a concrete proposition to test: a fixed language model can help build software that learns the rules as it goes.