Proposal
Add a TransformerBridge adapter for OuroForCausalLM (ByteDance Ouro / LoopLM), a Universal-Transformer-style looped model: one weight-tied layer stack iterated T times, plus a learned adaptive-exit halting gate.
Motivation
Ouro adds two surfaces no supported model has: weight-sharing across depth — the same L-layer stack is applied iteratively (default T=4), refining the hidden state in latent space; and a learned per-step exit/halting gate (ACT-style adaptive computation). Together they let researchers ask questions TL currently cannot pose: how an answer evolves/revises across loop iterations (cross-step logit lens, answer-flip analysis), what features drive halting (probe/patch the exit gate), and whether latent iterations are causally faithful vs post-hoc. It complements Huginn (RavenForCausalLM) by pairing the depth axis with adaptive-compute halting.
Gap scan (2026-06-25): ~15 models, ~288K downloads.
Scope note (medium-effort adapter)
Remote-code (auto_map to modeling_ouro), but the loop is a conventional block run in a for-loop — closer to a standard adapter than Huginn's variable-count recurrence. Remote-code loading is already supported (see openelm.py).
Pitch
Map the single shared stack iterated T times (expose each iteration's residual stream as a hookable step) plus the adaptive-exit gate head running parallel to the LM head.
- Claude Code users can scaffold with
/add-model-support ByteDance/Ouro-1.4B.
- Register at the four sites listed in contributing.md.
- Verify smallest-first:
ByteDance/Ouro-1.4B, then ByteDance/Ouro-2.6B-Thinking.
Additional context
Checklist
Proposal
Add a TransformerBridge adapter for
OuroForCausalLM(ByteDance Ouro / LoopLM), a Universal-Transformer-style looped model: one weight-tied layer stack iterated T times, plus a learned adaptive-exit halting gate.Motivation
Ouro adds two surfaces no supported model has: weight-sharing across depth — the same L-layer stack is applied iteratively (default T=4), refining the hidden state in latent space; and a learned per-step exit/halting gate (ACT-style adaptive computation). Together they let researchers ask questions TL currently cannot pose: how an answer evolves/revises across loop iterations (cross-step logit lens, answer-flip analysis), what features drive halting (probe/patch the exit gate), and whether latent iterations are causally faithful vs post-hoc. It complements Huginn (
RavenForCausalLM) by pairing the depth axis with adaptive-compute halting.Gap scan (2026-06-25): ~15 models, ~288K downloads.
Scope note (medium-effort adapter)
Remote-code (
auto_maptomodeling_ouro), but the loop is a conventional block run in afor-loop — closer to a standard adapter than Huginn's variable-count recurrence. Remote-code loading is already supported (see openelm.py).Pitch
Map the single shared stack iterated T times (expose each iteration's residual stream as a hookable step) plus the adaptive-exit gate head running parallel to the LM head.
/add-model-support ByteDance/Ouro-1.4B.ByteDance/Ouro-1.4B, thenByteDance/Ouro-2.6B-Thinking.Additional context
hf_scraperarchitecture-gaps pass (2026-06-25).Checklist