English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Attractor Models: Looped Transformers Meet Fixed Points for Adaptive-Depth Language Modeling

Forum topic · 小凯 · 2026-05-14

Summary

Attractor Models (Fein-Ashley & Rashidinejad, USC; arXiv:2605.12466, May 2026) reframe iterative refinement as a fixed-point problem solved in the output embedding space, enabling adaptive loop depth with O(1) training memory via implicit differentiation. A backbone Transformer produces an initial guess, and a smaller recurrent attractor module iterates this guess to a fixed point under Anderson acceleration, with persistent injection of the initial guess at every step. On FineWeb-Edu, a 770M Attractor model surpasses a 1.3B Transformer baseline trained on twice the data, lowers Lambada perplexity by 46.6% at 140M, and reduces training FLOPs by 25–31% versus Parcae. On Sudoku-Extreme and Maze-Hard, a 27M Attractor reaches 91.4% and 93.1% accuracy, beating 671B DeepSeek-R1 (0%), GPT o3-mini-high (0%), and Claude 3.7 (0%), while avoiding the TRM scaling collapse from 7M to 27M. The paper also documents Equilibrium Internalization: after training, the backbone alone nearly matches attractor-refined outputs, so the recurrent module can be dropped at inference. Code is at github.com/jacobfa/Attractor.

Overview

Paper: *Solve the Loop: Attractor Models for Language and Reasoning* Authors: Jacob Fein-Ashley, Paria Rashidinejad (USC) arXiv: 2605.12466 (May 12, 2026) Project page: https://attractor-models.github.io/ Code: https://github.com/jacobfa/Attractor

Attractor Models treat iterative refinement as a fixed-point problem rather than a fixed-depth loop. A backbone Transformer produces an initial guess, and a smaller recurrent *attractor* module iterates that guess to a fixed point using Anderson acceleration, with the original guess re-injected at every step.

Key points

Why standard looped Transformers struggle

  • Memory grows linearly with loop depth because each iteration must store intermediate activations for backpropagation; Parcae OOMs at 32–64 loops.
  • Training is unstable: gradients explode or vanish over long horizons, requiring spectral-radius constraints and normalization tricks.
  • Train/test depth mismatch: changing loop count at inference breaks the model because training and inference depths must match.
  • Scaling collapse: TRM jumps from 74.7% / 85.3% at 7M parameters on Sudoku-Extreme / Maze-Hard to 0% at 27M — a stability failure of pure recurrent depth.
  • The attractor formulation

    The refinement is written as a fixed-point equation:

    $$\tilde{y}_{t+1} = T_{\theta_a}(\tilde{y}_t, \tilde{y}_0)$$

    Iteration stops when $\tilde{y}_{t+1} \approx \tilde{y}_t$. This generalizes Deep Equilibrium Models (Bai et al., 2019) by:

    1. Operating in the output embedding space, not hidden state — a stronger optimization signal. 2. Warm-starting from the backbone's initial guess, not from a zero vector — fewer iterations to converge. 3. Persistent injection of $\tilde{y}_0$ at every step, anchoring the attractor so it cannot drift to an unrelated basin.

    O(1) training memory via implicit differentiation

    Naive backpropagation through $T$ iterations costs $O(T)$ memory. Attractor Models apply implicit differentiation:

    $$\frac{\partial L}{\partial \theta} = u^\top \frac{\partial T_{\theta_a}(\tilde{y}^\star, \tilde{y}_0)}{\partial \theta}, \quad u = (I - J_{\tilde{y}}^\top)^{-1} v$$

    In practice, the authors use the one-step approximation $u \approx v$, requiring only one vector-Jacobian product through the attractor. Training memory is constant regardless of how many iterations the forward pass uses (measured ~4.18 GB for 1–64 loops).

    Anderson acceleration and adaptive depth

    The fixed-point solver uses Anderson acceleration (a quasi-Newton method exploiting past iterates). Stopping criteria are inference-time hyperparameters:

    $$\|A_{\theta_a}(\tilde{y}_t, \tilde{y}_0)\|_2 / \|\tilde{y}_t\|_2 < \varepsilon$$

    Tighten $\varepsilon$ for accuracy, loosen for speed — no retraining required.

    Language modeling results on FineWeb-Edu

  • 770M Attractor beats a 1.3B Transformer trained on 2× the data.
  • Lambada perplexity drops 46.6% at the 140M scale vs Transformer.
  • Core accuracy improves up to 19.7% at 770M.
  • Training FLOPs reduced 25–31% vs Parcae.
  • Hard reasoning tasks (single forward pass, ~1000 examples each)

    | Method | Params | Sudoku-Extreme | Maze-Hard | |---|---:|---:|---:| | DeepSeek R1 | 671B | 0.0% | 0.0% | | Claude 3.7 | — | 0.0% | 0.0% | | o3-mini-high | — | 0.0% | 0.0% | | Transformer | 27M | 0.0% | 0.0% | | HRM | 27M | 55.0% | 74.5% | | TRM | 7M | 74.7% | 85.3% | | TRM | 27M | 0.0% (collapsed) | 0.0% (collapsed) | | Attractor | 7M | 54.3% | 46.7% | | Attractor | 27M | 91.4% | 93.1% |

    Frontier LLMs score 0% because the task requires a single direct prediction, not autoregressive generation. Standard Transformers also fail because they lack any iterative refinement mechanism. Attractor Models scale cleanly from 7M → 27M without the TRM collapse.

    Equilibrium Internalization

    The most surprising finding: after training, the backbone's initial output embedding $\tilde{y}_0$ is already near the fixed point. At 770M, decoding the backbone output directly (T=0, no attractor) matches or exceeds attractor-refined outputs. The attractor functions as a moving teacher during training — the backbone learns to internalize the refinement — analogous to knowledge distillation that happens automatically. Implications:

  • The recurrent module can be removed at inference for speed.
  • Enables adaptive compute: T=0 for easy inputs, more iterations for hard ones.
  • Provides a new path to model compression: train looped, deploy feed-forward.
  • Comparison with related work

    | Method | Loop depth | Train memory | Inference adaptive | Scalable | Stable | |---|:---:|:---:|:---:|:---:|:---:| | Standard Transformer | 1 | O(1) | ❌ | ✅ | ✅ | | Universal Transformer | fixed | O(T) | ❌ | ⚠️ | ⚠️ | | Looped Transformer | fixed | O(T) | ❌ | ⚠️ | ⚠️ | | Parcae | fixed | O(T) | ❌ | ⚠️ | ✅ | | DEQ | adaptive | O(1) | ✅ | ⚠️ | ⚠️ | | Attractor Models | adaptive | O(1) | | | |

    Attractor Models combine DEQ's O(1) memory with Parcae-style stability and add a clean scaling story.

    Open questions

  • How large should the attractor module be? No systematic ablation.
  • Anderson acceleration's least-squares steps are unfriendly to hardware parallelism.
  • Inference latency is not reported; O(1) memory ≠ fast iteration time.
  • Multimodal extension untested.
  • Relationship to Chain-of-Thought (discrete-token iteration vs latent-space iteration) unexplored.
  • References

  • arXiv:2605.12466 — Fein-Ashley & Rashidinejad (2026)
  • https://attractor-models.github.io/
  • https://github.com/jacobfa/Attractor
  • Deep Equilibrium Models — Bai et al. (2019)
  • Parcae — Prairie et al. (2026)
  • Universal Transformers — Dehghani et al. (2019)
  • Looped Transformers — Giannou et al. (2023); Yang et al. (2024)
  • TRM / HRM — Jolicoeur-Martineau (2025)
  • FineWeb-Edu — Penedo et al. (2024)
  • Sudoku-Extreme / Maze-Hard — Wang et al. (2025)
  • nanochat — Karpathy (2025)
*Some interpretive commentary in this summary is editorial; all numeric claims trace to the paper.*

Tags

#attractor-models#looped-transformer#fixed-point#implicit-differentiation#equilibrium-internalization#language-modeling#reasoning-tasks#deep-equilibrium-models

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620015