Per-Token Fixed-Point Convergence in Depth-Recurrent Transformers

• URL: http://arxiv.org/abs/2607.14427v1
• Date: Wed, 15 Jul 2026 23:38:44 GMT
• Title: Per-Token Fixed-Point Convergence in Depth-Recurrent Transformers
• Abstract summary: We show a training-free rule that halts each token once its output stabilizes attains uniform depth-8 quality at 4.94 average loops.<n>A linear router trained on convergence labels harvested from the same model requires nearly full depth and yields no reduction.<n>We report average depth as a FLOP proxy with a three-point wall-clock bracket rather than a realized speedup.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths. We ask what such a model actually computes per token, and measure it directly. On a 135M-class model trained on FineWeb-Edu, the recurrent state converges to a per-token fixed point: mean successive-output KL divergence falls from 3.9e-1 at the second loop to 8.5e-6 by the sixteenth, and per-token state change decays in step. Crucially, this convergence is not uniform across tokens. The median token converges by loop six, while approximately 10 percent of tokens continue to update at the training-mean depth of eight, and mean convergence depth is ordered by token type (whitespace shallowest, content words deepest). This per-token variation is the central object of the paper. We show it is directly readable and that reading it outperforms learning to predict it: a training-free rule that halts each token once its output stabilizes attains uniform depth-8 quality at 4.94 average loops (a 38 percent reduction in average depth) and matches uniform depth across the average-depth range, whereas a linear router trained on convergence labels harvested from the same model requires nearly full depth and yields no reduction. The elasticity that makes this possible reproduces here as background (validation loss decreases monotonically from 3.80 at one loop to 3.20 at eight and remains stable to 32 loops). We report average depth as a FLOP proxy with a three-point wall-clock bracket rather than a realized speedup, make no FLOP-matched parity claim, and note that the allocation results are established at a single scale and seed. The complete study runs on a single RTX 4090 in approximately 100 GPU-hours.

Related papers

• RecurTrace: Adaptive Latent Reasoning with Loop-Time Memory [25.81050020222202]
  Loop Memory Attention lets each looped layer attend to its own states from previous iterations along the loop-time axis.<n>RecurTrace achieves 56.9% accuracy with an average of 2.0 loops, exceeding the best fixed loop depth by 2.2 points at matched compute.
  arXiv  Detail & Related papers  (2026-09-03T05:29:24Z)
• On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability [57.03393882933515]
  We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and 51B parameters of n-gram embedding tables held off the accelerator.<n>On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs.
  arXiv  Detail & Related papers  (2026-08-31T06:35:07Z)
• Wiring Beats Blending: What Transfers Between Transformer Sizes -- and What Doesn't [0.0]
  Wecharacterize the 1.4B->410M conversion in the Pythia family end to end.<n>Weight residuals are statistically indistinguishable from noiseunder shuffle controls.<n>Compensationis a token-efficient, low-budget win rather than a universal one.
  arXiv  Detail & Related papers  (2026-08-03T19:44:08Z)
• DeepLoop: Depth Scaling for Looped Transformers [91.7922038545625]
  Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds.<n>We formalize this tied-depth effect through a first-order perturbation bound controlled by a visit-alignment coefficient.<n>The resulting method, textbfDeepLoop, keeps the Post-LN DeepNorm architecture and sets $= (2N)1/2$ and $=(8N)-1/2$ for unrolled depth.
  arXiv  Detail & Related papers  (2026-07-15T06:37:05Z)
• Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata [57.27430779838529]
  We present an exact and tractable algorithm for sampling from the constrained mean-field posterior under any constraint expressible as a finite automaton.<n>The approach guarantees constraint satisfaction by construction, supports both greedy and sampling-based decoding, and is compatible with parallel and block-wise decoding.<n> Empirical evaluations on Dream-7B and LLaDA-8 show substantial accuracy gains across various tasks.
  arXiv  Detail & Related papers  (2026-07-08T05:48:57Z)
• Learning to Solve Generative ODEs Beyond the Linear Span [50.13853710831612]
  We propose SpanLift, a lightweight neural solver that augments scalar-coefficient updates with a spatial residual operator.<n>Across pixel-space diffusion, latent flow matching, and precipitation nowcasting, SpanLift achieves state-of-the-art few-step sampling.
  arXiv  Detail & Related papers  (2026-06-07T15:22:13Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• CART: Context-Anchored Recurrent Transformer -- A Parameter-Efficient Architecture with Learned Stability [0.0]
  CART (Context-Anchored Recurrent Transformer) is a parameter-efficient language model that reuses a single shared core block R times across depth.<n>We evaluate CART on single consumer GPUs in two stages: a 64-configuration screen at 3,000 steps, then 36 configurations (P=6, R in 6,8,10, three seeds) trained for 30,500 steps (1B tokens)<n>Two patterns hold across widths d in 256,512,768,1024: prelude depth P dominates loop count R, and the Stage-1 ranking of R reverses at full training (R=6 becomes best at
  arXiv  Detail & Related papers  (2026-05-31T23:26:27Z)
• Universal Transformers Need Memory: Depth-State Trade-offs in Adaptive Recursive Reasoning [0.0]
  We study learned memory tokens as computational scratchpad for a single-block Universal Transformer (UT) on SudokuExtreme.<n>We find that memory tokens are empirically necessary: across all configurations tested, no configuration without memory tokens non-trivial performance.
  arXiv  Detail & Related papers  (2026-04-23T18:30:01Z)
• MeanFlow Transformers with Representation Autoencoders [71.45823902973349]
  MeanFlow (MF) is a diffusion-motivated generative model that enables efficient few-step generation by learning long jumps directly from noise to data.<n>We develop an efficient training and sampling scheme for MF in the latent space of a Representation Autoencoder (RAE)<n>We achieve a 1-step FID of 2.03, outperforming vanilla MF's 3.43, while reducing sampling GFLOPS by 38% and total training cost by 83% on ImageNet 256.
  arXiv  Detail & Related papers  (2025-11-17T06:17:08Z)
• Temporal Zoom Networks: Distance Regression and Continuous Depth for Efficient Action Localization [6.908972852063454]
  Temporal action localization requires both precise boundary detection and computational efficiency.<n>We address this through two complementary innovations: Boundary Distance Regression (BDR) and Adaptive Temporal Refinement (ATR)<n>On THUMOS14, our method achieves 56.5% mAP@0.7 with 151G FLOPs, using 36% fewer FLOPs than ActionFormer++ (55.7% mAP@0.7 at 235G)
  arXiv  Detail & Related papers  (2025-11-06T00:41:54Z)
• DeepPrune: Parallel Scaling without Inter-trace Redundancy [53.62015294143274]
  Over 80% of parallel reasoning traces yield identical final answers, representing substantial wasted computation.<n>We propose DeepPrune, a novel framework that enables efficient parallel scaling through dynamic pruning.<n>Our work establishes a new standard for efficient parallel reasoning, making high-performance reasoning more efficient.
  arXiv  Detail & Related papers  (2025-10-09T17:24:54Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.