Proposal
Add a TransformerBridge adapter for HrmTextForCausalLM (Sapient Intelligence HRM-Text), a hierarchical reasoning model with nested two-timescale depth recurrence.
Motivation
HRM is a third, distinct flavor of the depth-recurrence axis (alongside Huginn's variable-count loop and Ouro's halting loop): a slow high-level stack iterates over a fast low-level stack, with cross-stack additive state coupling. This gives interp researchers a clean substrate to separate slow "planning" from fast "computing" state, patch later refinement back into earlier steps, and watch attractor/fixed-point dynamics, none of which are expressible on a flat block list. There is direct mech-interp on the HRM architecture family (fixed-point / spurious-attractor / grokking analysis), though so far on the 27M puzzle HRM rather than HRM-Text specifically. HRM's claims are debated, which is itself a reason interp tooling on it is valuable.
Gap scan (2026-06-25): ~6 models, ~297K downloads.
Scope note (medium-effort adapter)
Native — model_type: hrm_text is supported in transformers >= 5.9.0 (model card: "ships native support for the hrm_text model class"). The repo currently pins transformers to 5.8.1, so picking this up needs a routine pin bump first. Verified recurrence fields: H_cycles=2, L_cycles=3 (plus L_bp_cycles=[0,3]); per-stack depth is num_hidden_layers (=16, "16 layers per stack"). The adapter maps the slow/fast recurrence onto the reused physical weights.
Pitch
Map the slow (H) and fast (L) stacks and their cross-stack coupling, exposing each as a hookable stream so the two timescales can be ablated and patched independently.
- Claude Code users can scaffold with
/add-model-support sapientinc/HRM-Text-1B.
- Register at the four sites listed in contributing.md.
- Verify smallest-first:
sapientinc/HRM-Text-1B, then LLM-OS-Models/KoHRM-Text-1.4B.
Additional context
Checklist
Proposal
Add a TransformerBridge adapter for
HrmTextForCausalLM(Sapient Intelligence HRM-Text), a hierarchical reasoning model with nested two-timescale depth recurrence.Motivation
HRM is a third, distinct flavor of the depth-recurrence axis (alongside Huginn's variable-count loop and Ouro's halting loop): a slow high-level stack iterates over a fast low-level stack, with cross-stack additive state coupling. This gives interp researchers a clean substrate to separate slow "planning" from fast "computing" state, patch later refinement back into earlier steps, and watch attractor/fixed-point dynamics, none of which are expressible on a flat block list. There is direct mech-interp on the HRM architecture family (fixed-point / spurious-attractor / grokking analysis), though so far on the 27M puzzle HRM rather than HRM-Text specifically. HRM's claims are debated, which is itself a reason interp tooling on it is valuable.
Gap scan (2026-06-25): ~6 models, ~297K downloads.
Scope note (medium-effort adapter)
Native —
model_type: hrm_textis supported intransformers >= 5.9.0(model card: "ships native support for the hrm_text model class"). The repo currently pinstransformersto 5.8.1, so picking this up needs a routine pin bump first. Verified recurrence fields:H_cycles=2,L_cycles=3(plusL_bp_cycles=[0,3]); per-stack depth isnum_hidden_layers(=16, "16 layers per stack"). The adapter maps the slow/fast recurrence onto the reused physical weights.Pitch
Map the slow (H) and fast (L) stacks and their cross-stack coupling, exposing each as a hookable stream so the two timescales can be ablated and patched independently.
/add-model-support sapientinc/HRM-Text-1B.sapientinc/HRM-Text-1B, thenLLM-OS-Models/KoHRM-Text-1.4B.Additional context
hf_scraperarchitecture-gaps pass (2026-06-25).Checklist