TYMED TYPER (TRM): The Senior Model is 7m Over the Deepseek-R1, Gemini 2.5 Pro, and O3-mini in consultation of the Arg-Agi 1 and AGI-AGI 2

Is the draft draft draft and repeatedly scratchpad Exterform largest autoregreate large autoregreate bigmy in Arc-Agi? Samsung Sait (Montreal) Issued A thin model repeat (TRM)-~ 7m-parameter recemble FICERER reports 44.6-45% the accuracy of the test in Arc-AGI-1 including 7.8-8% despite of- Arc-AGI-2The past results reported the models that are very hardened as Deepseek-R1, O3-Mini-High, and Gemini 2.5 Pro to the same social examination. TRM also improves puzzle benches Sudoku-Over (87.4%) including Maze-Hard (85.3%) over the previous Hierarchical consultant model (HRM, 27m params)While using a few parameters away from a simple recipe.

What's new new?

TRM removes HRM's AN-MODE Hierarchy and Fixed-Point Gradient Opceriment that allows a One small network that returns to “scratchpad” (z) as well as a Maximum Solution Current (Y):

• One of the only one small ordinary. Instead of the proposal of Hrm Module Hierarchy in one 2 network associated with the Scratchpad ¥ and the current solution to install 𝑦 Y. Model: Law: Refresh 𝑦 ← ← 𝑔 (𝑦, 𝑧) y ← g (y, z).
• Deeply targeted repetition. The → → Block Block Priorished 16 times With a deeper employment and the stigma used used during training (full prevention during testing). The signals are carried by steps with (Y, z) (y, z) (y, z).
• Full Backprop with LOOP. Unlike HRM's One-Step perfect (Fixed-Point) Gradient limitations, TRM Backpropagites through all recurring stepsThe research team that receives is important for normal performance.

Architactaly, the best setting of the arc / maze retains Ignorance; Small grids organized, the research team exchanges self-esteem with MLP-Mixer-Style Token Mixer. Little EMA (Exponential Average Average) above weight enforcement of training in restricted data. Net depth has been successfully created stun (eg. T = 3, n = 6) rather than enter the layers; On the face, Two layers Do better better than deep variations in the same compute.

Understanding the results

• Arc-AGI-1 / ARC-AGI-2 (two attempts): TRM-ATTN (7M): 44.6% / 7.8% vs HM (27m): 40.3% / 5.0%. Research Group – Report LLM Basenes: Deepseek-R1 (671b) 15.8% / 1.3%, O3-Mini-High 34.5% / 3.0%, Gemini 2.5 Pro 37.0% / 4.9%; More broke grok-4 entries is high (66.7-79.6% / 16-29.4%).
• Sudoku-Extreme (9 × 9, 1K test train): 87.4% With vs-free mixer vs HRM 55.0%.
• Maze-Hard (30 × 30): 85.3% vs hrr 74.5%.

These are direct predictions The models are trained from the beginning to small, healthy realities – not a few persevered. The arc remains a bookshot of books; Broadboard Leaderboard & Rules (eg arc-agi-prize limit at 85% set) are followed by the arc price for.

Why can a 7m model hit large llms in these activities?

1. Determined – Reviews instead of Token-by-Token: TRM writes a complete solution to make choices, and then it improves it with The chairs of hot variables against reducing the recognition of exposure from the default tests of formal exit.
2. Compute used in the thinking of the test period, not a parameter count: The active depth comes back back (depth of prepared ≈ T · (N + 1) · Background), investigators show better illustration of better illegal production than to add layers.
3. Hard selection is hard in grid thinking: In small grids organized (eg sudoku), careless mixing reduces the overall law and improves bias trade / difference; Self-maintaining is stored in large 30 × 30 grids.

Healed Key

• Building: ~ 7m-Param, the recurring solver 2 exchanging updates “Think” ←
• Results: Preface ~ 44.6-45% despite of- Arc-AGI-1 including ~ 7.8-8% despite of- Arc-AGI-2 (Try two)Exceeding several llms larger as compared to research paper comparison (eg Gemini 2.5 Pro, O3-Mini-High, Deepsek-R1) under the prescribed test protocol.
• Efficiency / Pattern: It shows that the computer's allocation of renewal update (depth of America) can strike parameters in the blue of the geometric, provides compact, from the Scracherch record with public published code.

This study shows ~ 7m-Parameter, the recurring solver that can renew up to 16 rounds team issuing the code in GitTub. Arc-Agi remains not fixed on a scale (a target 85% on ARC-AGI-2), therefore an offering of the property rather than successful thinking.

Look Technicine including GitHub page. Feel free to look our GITHUB page for tutorials, codes and letters of writing. Also, feel free to follow it Sane and don't forget to join ours 100K + ml subreddit Then sign up for Our newspaper. Wait! Do you with a telegram? Now you can join us with a telegram.

Asphazzaq is a Markteach Media Inc. According to a View Business and Developer, Asifi is committed to integrating a good social intelligence. His latest attempt is launched by the launch of the chemistrylife plan for an intelligence, MarktechPost, a devastating intimate practice of a machine learning and deep learning issues that are clearly and easily understood. The platform is adhering to more than two million moon visits, indicating its popularity between the audience.