Latent Phase-Shift Rollback: Inference-Time Error Correction via Residual Stream Monitoring and KV-Cache Steering

• URL: http://arxiv.org/abs/2604.18567v1
• Date: Mon, 20 Apr 2026 17:53:33 GMT
• Title: Latent Phase-Shift Rollback: Inference-Time Error Correction via Residual Stream Monitoring and KV-Cache Steering
• Abstract summary: Large language models commit unrecoverable reasoning errors mid-generation.<n>We introduce $textbfLatent Phase-Shift Rollback$ (LPSR)<n>At each generation step, we monitor the residual stream at a critical layer lcrit, detect abrupt directional reversals (phase shifts) via a cosine-similarity $+$ entropy dual gate.<n>No fine-tuning, gradient computation, or additional forward passes are required.
• Score: 4.032680910442999
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models frequently commit unrecoverable reasoning errors mid-generation: once a wrong step is taken, subsequent tokens compound the mistake rather than correct it. We introduce $\textbf{Latent Phase-Shift Rollback}$ (LPSR): at each generation step, we monitor the residual stream at a critical layer lcrit, detect abrupt directional reversals (phase shifts) via a cosine-similarity $+$ entropy dual gate, and respond by rolling back the KV-cache and injecting a pre-computed steering vector. No fine-tuning, gradient computation, or additional forward passes are required. LPSR achieves $\mathbf{44.0\%}$ on MATH-500 with an 8B model versus $28.8\%$ for standard AR ($+15.2$ pp; McNemar $χ^2 = 66.96$, $p < 10^{-15}$). Critically, prompted self-correction, the most natural inference-time baseline, scores only $19.8\%$, below standard AR; LPSR exceeds it by $+24.2$ pp ($χ^2 = 89.4$, $p \approx 0$). LPSR also outperforms Best-of-16 ($+7.8$ pp) at $5.4\times$ lower token cost, and surpasses a standard 70B model ($35.2\%$) with $8.75\times$ fewer parameters at ${\sim}3\times$ the token budget. A 32-layer sweep reveals a novel \textbf{detection-correction dissociation}: error-detection AUC peaks at layer~14 ($0.718$) but task accuracy peaks at layer~16 ($44.0\%$ vs.\ $29.2\%$), demonstrating that optimal monitoring depth differs for detection and correction.

Related papers

• Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors [51.56484100374058]
  Autoregressive models accumulate error over long rollouts, yet at deployment there is no ground truth to measure it against.<n>We train a single conditional latent diffusion model that steps a dynamical system forward or backward in time via a direction flag.<n>We show that this bidirectionality supplies a measurement-free test-time error signal.
  arXiv  Detail & Related papers  (2026-08-01T13:49:46Z)
• ForgettingOT: Certified Speculative Batching from Sinkhorn's Projective Forgetting [0.0]
  We show that the eigenmode of the fixed-point Jacobian $J_tstar=QP=Pstar P$ controls the strict correction tail.<n>ForgettingOT turns this nonlinear Perron--Frobenius fact into a certified executor for streams of related Sinkhorn problems.
  arXiv  Detail & Related papers  (2026-07-27T17:59:36Z)
• Calibrated e-CUSUM Decoding for Quantized Reasoning Models: Why Token Log-Probability Is the Wrong Observable for Decoding Monitors [0.6999740786886536]
  Low-bit quantization makes small reasoning models inexpensive to deploy but can degrade their chains of thought.<n>We show that the centered token log-probability increment $log p(w_t)+H_t$ is the wrong observable for this purpose.<n>We introduce a training-free decoding controller that combines a degeneration-aware alarm score fusing token uncertainty with explicit repetition.
  arXiv  Detail & Related papers  (2026-07-13T09:34:36Z)
• An End-to-End Hybrid Quantum--Classical Sampling Workflow for Discrete Markov Random Fields: A Reproducible Case Study [0.0]
  We study amplitude-encoded i.i.d. sampling for small discrete Markov random fields (MRFs)<n>This removes quantum exponential speedup but allows a clean comparison against classical MCMC based on independent circuit samples.<n>A matched-budget VQC vs. MPS comparison shows VQC fidelities fall far below MPS: $(F_mathrmVQC, F_mathrmMPS) = (0.31, 0.96), (0.21, 0.96), (0.17, 0.88)$ at compressions $10.7times$, $34.1times$, and $11
  arXiv  Detail & Related papers  (2026-07-10T18:33:14Z)
• Pruning Deep Neural Networks via the Marchenko--Pastur Distribution [0.0]
  We study a Marchenko--Pastur (MP) random-matrix approach to pruning deep neural networks with very small post-pruning fine-tuning budgets.<n>The main practical contribution is accuracy retention under short calibration and fine-tuning schedules.
  arXiv  Detail & Related papers  (2026-05-23T19:44:02Z)
• Hidden-State Privacy Has an Empty Middle [51.56484100374058]
  We prove a complementary Fisher-ball lower bound: every full-rank Gaussian release at $O(1)$ Fisher utility admits a direction whose Mahalanobis signal grows linearly in hidden width.<n>A split-memory transformer trained from scratch reaches $G_mathrmMah in [20, 33]$ at 90M and maintains a $6$--$24times$ advantage over same-budget GPT baselines from 30M to 1B at a fixed-token language loss penalty.
  arXiv  Detail & Related papers  (2026-05-21T20:12:09Z)
• Online Covariance Estimation in Averaged SGD: Improved Batch-Mean Rates and Minimax Optimality via Trajectory Regression [12.805268849262243]
  We study online covariance matrix estimation for Polyak-Ruppert averaged gradient descent (SGD)<n>The construction reveals that the bottleneck is the sublinear accumulation of information about the Hessian from the SGD drift.
  arXiv  Detail & Related papers  (2026-04-12T20:49:33Z)
• dynActivation: A Trainable Activation Family for Adaptive Nonlinearity [0.0]
  dynActivation variants tend to linearize deep layers while maintaining high performance.<n>A new proposed dynActGLU-variant achieves a $10.3%$ relative perplexity reduction over SwiGLU at 5620 steps (4.047 vs. 4.514), though the gap vanishes at 34300 steps.
  arXiv  Detail & Related papers  (2026-03-23T16:18:28Z)
• INC: An Indirect Neural Corrector for Auto-Regressive Hybrid PDE Solvers [61.84396402100827]
  We propose the Indirect Neural Corrector ($mathrmINC$), which integrates learned corrections into the governing equations.<n>$mathrmINC$ reduces the error amplification on the order of $t-1 + L$, where $t$ is the timestep and $L$ the Lipschitz constant.<n>We test $mathrmINC$ in extensive benchmarks, covering numerous differentiable solvers, neural backbones, and test cases ranging from a 1D chaotic system to 3D turbulence.
  arXiv  Detail & Related papers  (2025-11-16T20:14:28Z)
• CarBoN: Calibrated Best-of-N Sampling Improves Test-time Reasoning [62.56541355300587]
  We introduce a general test-time calibration framework that adaptively modifies the model toward high-reward reasoning paths.<n>Within this framework, we propose CarBoN, a two-phase method that first explores the solution space and then learns a calibration of the logits.<n>Experiments on MATH-500 and AIME-2024 show that CarBoN improves efficiency, with up to $4times$ fewer rollouts to reach the same accuracy.
  arXiv  Detail & Related papers  (2025-10-17T14:04:37Z)
• Numerical Fragility in Transformers: A Layer-wise Theory for Explaining, Forecasting, and Mitigating Instability [0.0]
  We give a first-order, module-wise theory that predicts when and where errors grow.<n>For self-attention we derive a per-layer bound that factorizes into three interpretable diagnostics.<n>We also introduce a precision- and width-aware LayerNorm indicator $rho_rm LN$ with a matching first-order bound.
  arXiv  Detail & Related papers  (2025-10-17T01:03:02Z)
• Variance-Dependent Regret Lower Bounds for Contextual Bandits [65.93854043353328]
  Variance-dependent regret bounds for linear contextual bandits, which improve upon the classical $tildeO(dsqrtK)$ regret bound to $tildeO(dsqrtsum_k=1Ksigma_k2)$.
  arXiv  Detail & Related papers  (2025-03-15T07:09:36Z)
• Simple ReFlow: Improved Techniques for Fast Flow Models [68.32300636049008]
  Diffusion and flow-matching models achieve remarkable generative performance but at the cost of many sampling steps. We propose seven improvements for training dynamics, learning and inference. We achieve state-of-the-art FID scores (without / with guidance, resp.) for fast generation via neural ODEs.
  arXiv  Detail & Related papers  (2024-10-10T11:00:55Z)
• Robust Nonparametric Regression under Poisoning Attack [13.470899588917716]
  An adversarial attacker can modify the values of up to $q$ samples from a training dataset of size $N$. Our initial solution is an M-estimator based on Huber loss minimization. The final estimate is nearly minimax optimal for arbitrary $q$, up to a $ln N$ factor.
  arXiv  Detail & Related papers  (2023-05-26T09:33:17Z)
• Horizon-Free and Variance-Dependent Reinforcement Learning for Latent Markov Decision Processes [62.90204655228324]
  We study regret minimization for reinforcement learning (RL) in Latent Markov Decision Processes (LMDPs) with context in hindsight. We design a novel model-based algorithmic framework which can be instantiated with both a model-optimistic and a value-optimistic solver.
  arXiv  Detail & Related papers  (2022-10-20T21:32:01Z)
• Optimal Robust Linear Regression in Nearly Linear Time [97.11565882347772]
  We study the problem of high-dimensional robust linear regression where a learner is given access to $n$ samples from the generative model $Y = langle X,w* rangle + epsilon$ We propose estimators for this problem under two settings: (i) $X$ is L4-L2 hypercontractive, $mathbbE [XXtop]$ has bounded condition number and $epsilon$ has bounded variance and (ii) $X$ is sub-Gaussian with identity second moment and $epsilon$ is
  arXiv  Detail & Related papers  (2020-07-16T06:44:44Z)

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.