Instruction Tuning Changes How Upstream State Conditions Late Readout: A Cross-Patching Diagnostic

• URL: http://arxiv.org/abs/2605.07284v1
• Date: Fri, 08 May 2026 05:47:10 GMT
• Title: Instruction Tuning Changes How Upstream State Conditions Late Readout: A Cross-Patching Diagnostic
• Abstract summary: We ask how earlier and the late stack cooperate to turn differences into next-token margins.<n>We cross each model's earlier-layer state with each model's late stack.<n>Forced-token scoring shows that the local token choice can change later exact-answer success.
• Score: 9.8812664524155
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent interpretability work has identified model-internal handles on post-trained behavior, including refusal directions, assistant/persona axes, and sparse chat-tuning features. These results localize where behaviors can be read out or controlled, often in middle-to-late layers. We ask how earlier computation and the late stack cooperate to turn those differences into next-token margins. To test this, we introduce first-divergence cross-patching: at the first token where pretrained base (PT) and instruction-tuned (IT) checkpoints disagree, we cross each model's earlier-layer state with each model's late stack. The diagnostic separates training recipes: same-base instruction-following descendants show late effects that depend on their own earlier-layer state, while OpenMath2 math-domain SFT and controlled code/biomed CPT controls with verified domain learning do not; for OpenMath2, the late effect is already largely portable from base earlier-layer state. Across five dense families (4B-32B), the IT late stack adds +0.76 logits from PT upstream and +2.44 from IT upstream, giving a +1.68 interaction that is positive in every family. Thus the late stack has a real PT-upstream effect, but its larger effect in the IT checkpoint appears only when it reads its own post-trained upstream state. Sparse features in final MLP layers partially mediate the effect and are driven by upstream patches, supporting a handoff from earlier state to final-layer feature activation to IT-token margin. Forced-token scoring shows that the local token choice can change later exact-answer success. Operationally, paired-checkpoint studies that localize a difference to late layers should test whether it survives under the other checkpoint's upstream state before treating the late stack as self-contained.

Related papers

• Encoded Early, Used Late: Where Transformers Begin to Act on an Inferred Partner's Expertise [5.269026412088572]
  We show that a transformer can make an attribute linearly decodable in its residual stream at a depth where that attribute does not yet influence the output.<n>We ask whether it also holds for an attribute the model must infer gradually over a conversation, namely how expert its dialogue partner is.
  arXiv  Detail & Related papers  (2026-09-07T07:41:57Z)
• Behaviorally Effective LoRA Writes Are Sparse and Structured [0.0]
  Low-rank adaptation fixes the rank of the update, but it does not identify which parts of a trained write actually carry behavior.<n>We show that behaviorally effective LoRA writes are sparse, structured, and far more concentrated than the raw low-rank parameterization suggests.
  arXiv  Detail & Related papers  (2026-09-01T15:09:43Z)
• Towards Evaluating Data Priors for Tabular Foundation Models [74.75794086564035]
  We implement a unified interface for publicly available priors and priors constructed from real datasets.<n>We generate training tasks from each prior, train the same model architecture under a fixed training protocol, and evaluate the resulting models on shared downstream classification tasks.<n>Our results show that different priors favor different downstream behaviors, with some achieving stronger absolute performance and others exhibiting more consistent relative rankings across datasets.
  arXiv  Detail & Related papers  (2026-06-28T07:12:45Z)
• From Patches to Trajectories: Privileged Process Supervision for Software-Engineering Agents [56.31499185764872]
  Supervised fine-tuning (SFT) on long teacher trajectories is the dominant way to instill investigation and reasoning in open software-engineering (SWE) agents.<n>We propose Patches-to-Trajectories (P2T), which uses $pstar$ as privileged information during curation and formulates trajectory construction as bi-objective optimization over per-step effectiveness and trajectory length.
  arXiv  Detail & Related papers  (2026-05-21T04:54:55Z)
• The Convergence Gap: Instruction-Tuned Language Models Stabilize Later in the Forward Pass [9.8812664524155]
  Final outputs hide when a checkpoint commits to its next-token prediction.<n>We introduce the convergence gap, a model-diffing diagnostic that decodes each layer's next-token distribution and measures its distance to the model's own final distribution.
  arXiv  Detail & Related papers  (2026-05-08T05:45:52Z)
• CLUE: Non-parametric Verification from Experience via Hidden-State Clustering [64.50919789875233]
  We show that correctness of a solution is encoded as a geometrically separable signature within the trajectory of hidden activations.<n>ClUE consistently outperforms LLM-as-a-judge baselines and matches or exceeds modern confidence-based methods in reranking candidates.
  arXiv  Detail & Related papers  (2025-10-02T02:14:33Z)
• Pipeline Parallelism is All You Need for Optimized Early-Exit Based Self-Speculative Decoding [73.67253077506672]
  Large language models (LLMs) deliver impressive generation quality, but incur very high inference cost.<n>Early-exit based self-speculative decoding (EESD) has emerged to mitigate this cost.<n>We propose Pipeline-Parallel Self-Speculative Decoding (PPSD) that fully pipelines the draft and verification work.
  arXiv  Detail & Related papers  (2025-09-19T04:51:41Z)
• Learning effective pruning at initialization from iterative pruning [15.842658282636876]
  We present an end-to-end neural network-based PaI method to reduce training costs. Our approach outperforms existing methods in high-sparsity settings. As the first neural network-based PaI method, we conduct extensive experiments to validate the factors influencing this approach.
  arXiv  Detail & Related papers  (2024-08-27T03:17:52Z)
• Dual-Path Adversarial Lifting for Domain Shift Correction in Online Test-time Adaptation [59.18151483767509]
  We introduce a dual-path token lifting for domain shift correction in test time adaptation. We then perform dual-path lifting with interleaved token prediction and update between the path of domain shift tokens and the path of class tokens. Experimental results on the benchmark datasets demonstrate that our proposed method significantly improves the online fully test-time domain adaptation performance.
  arXiv  Detail & Related papers  (2024-08-26T02:33:47Z)
• LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding [13.747101397628887]
  We present an end-to-end solution to speed-up inference of large language models (LLMs) We apply layer dropout, with low dropout rates for earlier layers and higher dropout rates for later layers, and an early exit loss where all transformer layers share the same exit. We show that this training recipe increases the accuracy of early exit at earlier layers, without adding any auxiliary layers or modules to the model.
  arXiv  Detail & Related papers  (2024-04-25T16:20:23Z)
• Less Peaky and More Accurate CTC Forced Alignment by Label Priors [57.48450905027108]
  Connectionist temporal classification (CTC) models are known to have peaky output distributions. This paper aims at alleviating the peaky behavior for CTC and improve its suitability for forced alignment generation. Our CTC model produces less peaky posteriors and is able to more accurately predict the offset of the tokens besides their onset.
  arXiv  Detail & Related papers  (2024-04-22T17:40:08Z)
• Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization [39.14048972373775]
  Test-time domain adaptation aims to adapt the model trained on source domains to unseen target domains using a few unlabeled images. Previous works normally update the whole network naively without explicitly decoupling the knowledge between label and domain. We propose to reduce such learning interference and elevate the domain knowledge learning by only manipulating the BN layer.
  arXiv  Detail & Related papers  (2023-12-15T19:22:21Z)
• REGTR: End-to-end Point Cloud Correspondences with Transformers [79.52112840465558]
  We conjecture that attention mechanisms can replace the role of explicit feature matching and RANSAC. We propose an end-to-end framework to directly predict the final set of correspondences. Our approach achieves state-of-the-art performance on 3DMatch and ModelNet benchmarks.
  arXiv  Detail & Related papers  (2022-03-28T06:01:00Z)

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.