| # TRM-text |
|
|
| TRM-text is an attention-free language model based on a Tiny Recursive Model (TRM) architecture. |
|
|
| Unlike Transformer-based language models, TRM-text removes self-attention entirely and replaces it with recursive computation built from causal dilated depthwise convolutions, hierarchical latent refinement, and parameter reuse. |
|
|
| The goal of TRM-text is to investigate whether recursive neural computation can provide competitive language modeling performance while dramatically reducing computational cost. |
|
|
| --- |
|
|
| # Overview |
|
|
| TRM-text explores a different scaling path from Transformers. |
|
|
| Instead of increasing attention heads and context interactions, TRM-text repeatedly refines hidden representations using a hierarchy of recursive processing blocks. |
|
|
| Key properties: |
|
|
| * Attention-free |
| * Autoregressive language modeling |
| * Recursive computation |
| * Hierarchical latent refinement |
| * RoPE positional encoding |
| * Dilated depthwise convolution mixer |
| * Hugging Face compatible |
| * safetensors support |
|
|
| --- |
|
|
| # Architecture |
|
|
| ```text |
| Tokens |
| │ |
| ▼ |
| Embedding |
| │ |
| ▼ |
| Low-Level TRM |
| │ |
| ▼ |
| Mid-Level TRM |
| │ |
| ▼ |
| High-Level TRM |
| │ |
| ▼ |
| Recursive Feedback |
| │ |
| ▼ |
| LM Head |
| ``` |
|
|
| Each recursive block contains: |
|
|
| * RMSNorm |
| * RoPE |
| * Causal Dilated Depthwise Convolution |
| * SwiGLU Feed Forward Network |
| * Residual Recurrence |
|
|
| No self-attention layers are used. |
|
|
| --- |
|
|
| # Model Configuration |
|
|
| Current release: |
|
|
| ```text |
| Parameters: ~15M |
| |
| dim = 256 |
| hidden_dim = 512 |
| |
| low_steps = 6 |
| mid_steps = 3 |
| high_steps = 2 |
| |
| cycles = 3 |
| |
| kernel_size = 5 |
| |
| low_dilations = [1,2,4,8] |
| mid_dilations = [2,4,8,16] |
| high_dilations = [4,8,16,32] |
| ``` |
|
|
| --- |
|
|
| # Compute Efficiency |
|
|
| Relative training cost: |
|
|
| | Architecture | Relative Cost | |
| | ------------ | ------------: | |
| | Transformer | 1200 | |
| | HRM | 100 | |
| | TRM-text | 1 | |
|
|
| These values represent relative compute requirements under the experimental scaling assumptions used during development. |
|
|
| The objective of TRM-text is to maximize efficiency through: |
|
|
| * parameter reuse |
| * recursive computation |
| * hierarchical refinement |
| * elimination of attention operations |
|
|
| --- |
|
|
| # Training |
|
|
| ## Base Pretraining |
|
|
| Dataset: |
|
|
| ```text |
| FineWeb Sample-10BT |
| ``` |
|
|
| Tokenizer: |
|
|
| ```text |
| GPT-2 BPE |
| ``` |
|
|
| Objective: |
|
|
| ```text |
| Causal Language Modeling |
| ``` |
|
|
| --- |
|
|
| # Instruction Tuning |
|
|
| Dataset: |
|
|
| ```text |
| tatsu-lab/alpaca |
| ``` |
|
|
| Format: |
|
|
| ```text |
| ### Instruction: |
| ... |
| |
| ### Response: |
| ... |
| ``` |
|
|
| --- |
|
|
| # Loading |
|
|
| ```python |
| from transformers import AutoTokenizer |
| from transformers import AutoModelForCausalLM |
| |
| tokenizer = AutoTokenizer.from_pretrained( |
| "summerMC/TRM-text", |
| trust_remote_code=True |
| ) |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| "summerMC/TRM-text", |
| trust_remote_code=True |
| ) |
| ``` |
|
|
| --- |
|
|
| # Inference |
|
|
| ```python |
| prompt = """ |
| ### Instruction: |
| Explain artificial intelligence in simple terms. |
| |
| ### Response: |
| """ |
| |
| inputs = tokenizer( |
| prompt, |
| return_tensors="pt" |
| ) |
| |
| outputs = model.generate( |
| **inputs, |
| max_new_tokens=128, |
| do_sample=True, |
| temperature=0.7, |
| top_k=40 |
| ) |
| |
| print( |
| tokenizer.decode( |
| outputs[0], |
| skip_special_tokens=True |
| ) |
| ) |
| ``` |
|
|
| --- |
|
|
| # Research Motivation |
|
|
| TRM-text investigates whether recursive neural systems can replace attention mechanisms in language modeling. |
|
|
| Research directions: |
|
|
| * recursive reasoning |
| * hierarchical computation |
| * efficient language models |
| * attention-free architectures |
| * low-cost scaling laws |
|
|
| --- |
|
|
| # Limitations |
|
|
| Current checkpoint is experimental. |
|
|
| Known limitations: |
|
|
| * small parameter count |
| * limited instruction tuning |
| * lower capability than modern frontier models |
| * research-focused implementation |
| * benchmark coverage still limited |
|
|
| --- |
|
|
| # Intended Use |
|
|
| TRM-text is intended for: |
|
|
| * language model research |
| * efficient architecture experimentation |
| * recursive computation studies |
| * attention-free modeling research |
|
|
| Not intended for: |
|
|
| * safety-critical systems |
| * medical decision making |
| * legal advice |
| * financial advice |
|
|
| --- |
|
|
| # License |
|
|
| Apache-2.0 |
|
|
| --- |
|
|
| # Citation |
|
|
| ```bibtex |
| @software{trm_text_2026, |
| title={TRM-text: Attention-Free Recursive Language Modeling}, |
| author={summerMC}, |
| year={2026}, |
| url={https://huggingface.co/summerMC/TRM-text} |
| } |
| ``` |
|
|