Dense Supervision Is Not Enough: The Readout Blind Spot in Looped Language Models

• URL: http://arxiv.org/abs/2606.24898v1
• Date: Fri, 12 Jun 2026 13:23:18 GMT
• Title: Dense Supervision Is Not Enough: The Readout Blind Spot in Looped Language Models
• Abstract summary: We show that dense per-loop cross-entropy controls the variables exposed by the readout, not every variable active in the recurrent transition.<n>In 44M and 129M looped transformers without inter-loop normalization, per-loop cross-entropy through RMSNorm readouts still drives final hidden-state norms into the thousands or tens of thousands.
• Score: 5.903857251549266
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Looped language models turn hidden states into runtime state: each state is decoded for prediction and fed back into future computation. This creates a basic supervision question: which state variables does cross-entropy actually control? We show that dense per-loop cross-entropy controls the variables exposed by the readout, not every variable active in the recurrent transition. Hidden-state scale gives a concrete failure mode. Scale-invariant readouts such as RMSNorm and LayerNorm hide radial scale from the immediate cross-entropy loss, while pre-norm residual recurrence continues to carry and update that same scale. Thus per-loop loss can make early exits usable without controlling recurrent scale. In 44M and 129M looped transformers without inter-loop normalization, per-loop cross-entropy through RMSNorm readouts still drives final hidden-state norms into the thousands or tens of thousands. Scale-visible readouts and explicit norm penalties keep norms in the tens, and scale-removing recurrence is the complementary architectural fix. The resulting design rule is simple: dense supervision trains exits; recurrent scale control requires either making scale visible to a loss or removing it from the loop. Consistent with this rule, scale-controlled variants achieve lower perplexity at matched inference-depth operating points in our variable-depth benchmarks.

Related papers

• 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)
• Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models [75.80275843320511]
  We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed.<n>We propose Hidden Decoding, a sequence-length scaling method applied during continued pretraining.
  arXiv  Detail & Related papers  (2026-07-09T07:37:59Z)
• Grokking Is Conditional and Fragile: A Fully-Tractable, Multi-Seed Study at 12K Parameters [0.0]
  We study grokking as a multi-seed rate rather than a single outcome.<n>In this fully-tractable regime grokking is a conditional, fragile phase transition.
  arXiv  Detail & Related papers  (2026-07-06T13:58:31Z)
• Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping [56.14767235650558]
  Looped Transformers, which repeatedly apply a shared transformer block, are an architecturally natural fit for variable-length computation tasks.<n>We trace this variance to the spurious correlation in simple algorithmic tasks between sequence length and number of loops.<n>Our work suggests that "when to stop" should be treated as a training design choice, not merely an inference-time-allocation rule.
  arXiv  Detail & Related papers  (2026-06-29T08:58:09Z)
• What Does the Weight Norm Control in Grokking? Logit-Scale Mediation under Cross-Entropy [0.0]
  We ask what the norm actually controls.<n>We slide the grokking delay across its entire norm-induced range under cross-entropy.<n>Across a grid of norms and temperatures the delay collapses onto the logit scale alone (R2 = 0.97), with the norm adding 1-2% beyond it.
  arXiv  Detail & Related papers  (2026-06-16T20:16:18Z)
• Noise-Driven Escape from Metastable Phases explains Grokking in Deep Neural Networks [0.0]
  We show that deep neural networks (DNNs) exhibit first order phase transitions under variations of the L2 regularization strength.<n>We reproduce grokking-like delayed convergence across two orders of magnitude in escape time.<n>Our work suggests that the number of metastable states equals the number of learnable features.
  arXiv  Detail & Related papers  (2026-06-15T11:22:16Z)
• Latent Recurrent Transformer: Architecture Exploration, Training Strategies, and Scaling Behavior [107.2098567818173]
  Latent Recurrent Transformer (LRT) is a lightweight augmentation of autoregressive transformers.<n>LRT reuses a high-level source-layer hidden state from the previous token as recurrent memory for the next token.
  arXiv  Detail & Related papers  (2026-05-26T10:10:26Z)
• Silent Collapse in Recursive Learning Systems [24.578102943388256]
  We propose a lightweight metacognitive loop that monitors trajectory statistics and estimates a slow-timescale trust variable.<n>MTR provides early warning and actively prevents silent collapse without requiring access to pristine real data.
  arXiv  Detail & Related papers  (2026-05-14T08:59:26Z)
• Score-Repellent Monte Carlo: Toward Efficient Non-Markovian Sampler with Constant Memory in General State Spaces [48.37754141688874]
  We propose a framework that summarizes history by a running average of score evaluations in $Rd$, where $d$ is the dimension of the score and state representation.<n>This history is converted into a surrogate target through an exponential score tilt, indexed with $$ that represents the strength of repellence in controlling the magnitude of the history-based repulsion.<n>Experiments on continuous targets and discrete energy-based models demonstrate improved estimator variance and mode coverage, while retaining $O(d)$ memory usage and modest per-it overhead.
  arXiv  Detail & Related papers  (2026-04-24T18:39:50Z)
• Gated Removal of Normalization in Transformers Enables Stable Training and Efficient Inference [2.1665689529884697]
  We introduce TaperNorm, a drop-in replacement for RMSNorm/LayerNorm that behaves exactly like the standard normalizer early in training.<n>A single global gate is held at $g=1$ during gate warmup, used to calibrate the scaling branch via EMAs, and then cosine-decayed to $g=0$.<n>On an efficiency microbenchmark, folding internal scalings yields up to $1.22times$ higher throughput in last-token logits mode.
  arXiv  Detail & Related papers  (2026-02-11T01:40:34Z)
• Robust Layerwise Scaling Rules by Proper Weight Decay Tuning [50.11170157029911]
  In modern scale-invariant architectures, training quickly enters an degrading-governed steady state.<n>We introduce a weight-decay scaling rule for AdamW that preserves sublayer gain across widths.<n>Our results extend $mu$P beyond the near-init regime by explicitly controlling the steady-state scales set by parameters.
  arXiv  Detail & Related papers  (2025-10-17T02:58:35Z)
• TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning [79.59753528758361]
  We present TrajBooster, a cross-embodiment framework that leverages abundant wheeled-humanoid data to boost bipedal VLA.<n>Our key idea is to use end-effector trajectories as a morphology-agnostic interface.<n>Results show that TrajBooster allows existing wheeled-humanoid data to efficiently strengthen bipedal humanoid VLA performance.
  arXiv  Detail & Related papers  (2025-09-15T12:25:39Z)
• Multistep Inverse Is Not All You Need [87.62730694973696]
  In real-world control settings, the observation space is often unnecessarily high-dimensional and subject to time-correlated noise. It is therefore desirable to learn an encoder to map the observation space to a simpler space of control-relevant variables. We propose a new algorithm, ACDF, which combines multistep-inverse prediction with a latent forward model.
  arXiv  Detail & Related papers  (2024-03-18T16:36:01Z)

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.