The paper investigates in-context binding errors in language models, showing that a linear probe can recover correct entity bindings from frozen hidden states even when the model outputs incorrect bindings. Across 16 checkpoints, the probe’s accuracy on failure cases surpasses a baseline by about 0.196, and a probe‑based score improves failure detection over the model’s confidence by 0.079 AUROC. Steering the residual stream toward the probe‑decoded binding further boosts accuracy by an average of 0.168 across eight models.
By Manas Venkata Sai Ravulapalli, Samrath Singh Chadha, Abhinav M. Hari
The paper investigates whether the latent states reconstructed during inference match those stored during training in temporal models, focusing on State Discovery in World Machine. By comparing stored state tables with rollouts from the best‑validation checkpoint and the last‑epoch weights, the authors find that reloading the checkpoint used to write the table eliminates 90–91% of the mean per‑dimension Wasserstein‑1 distance, explaining most of the observed state gap. Causal distillation and frozen SD updates fail to meet their criteria under either checkpoint.
By Weijia Han, Lisha Qu
arXiv:2609.08618v1 Announce Type: new
Abstract: Benchmark scores describe what a checkpoint can do now, but they do not determine how it will respond to the next training episode. We measure this mis...
By Zhongxuan Liu, Sicheng Zhou, Hongzhi Wang
Temporal models that maintain per-sequence latent states during training must reconstruct those states when the stored table is unavailable at inference. This creates a basic consistency question: do...
The paper introduces the concept of intervention fidelity in latent world models, measuring whether a model’s open‑loop transitions align with actual environment interventions. Experiments on TD‑MPC2, Cheetah, and DreamerV3 show that high reward fit does not guarantee fidelity, and that self‑supervised models can outperform task‑anchored ones in preserving intervention effects. The authors propose a capture‑gated audit to localize failures and argue that fidelity must be directly audited on the model’s native interface.
By Donna Vakalis
arXiv:2609.36569v1 Announce Type: cross
Abstract: Checkpoint selection is a routine decision in supervised fine-tuning (SFT): training produces multiple checkpoints, but only one is retained. Yet fix...
By Yupeng Chang, Wenxuan Zhang, Yuan Wu
The paper introduces entry-state sharpening, a data‑free pre‑training step that prepares a language model’s checkpoint in a sharper, lower‑entropy state before test‑time reinforcement learning (TTRL). By reducing policy entropy, the model can more efficiently use its limited adaptation budget, leading to higher endpoint conversion efficiency across tasks such as MATH, GPQA, and AMC. Experiments with different data‑free objectives (e.g., R‑Zero vs. SPIRAL) demonstrate that the choice of pre‑training objective strongly influences the checkpoint’s readiness for TTRL, and a label‑free self‑distillation intervention can further sharpen the entry state.
By Zhanming Zhang, Vinoth Selvendran
arXiv:2607. 20436v1 Announce Type: cross Abstract: Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption.
By Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz
arXiv:2609.01048v1 Announce Type: cross
Abstract: Across the full Pythia suite (160M-12B, eight checkpoints, four task families), a linear probe can read a target variable from the residual stream as...
By Xining Xun
The paper introduces a benchmark for predicting a program’s final output and checkpoint state using only its source code and input, extending CRUXEval-style tasks with both short‑ and long‑trace inputs and checkpoints inside and after loops. It evaluates 400 cases from 371 Python and C++ programs across seven settings from four model families, achieving up to 93.0% accuracy on short‑trace final output and 77.0% on long‑trace final output, with reasoning‑enabled settings outperforming non‑reasoning ones by 33.1 to 55.2 percentage points. The study also shows that longer‑trace inputs can change correctness outcomes and that the benchmark reveals errors that short‑output scores alone would miss.
By Xiaohong Chen, David Bucur, Chenglong Ma, Yi Zhang, Lingming Zhang, Sriram Vishwanath, Grigore Rosu
The paper introduces a diagnostic framework for long‑horizon security LLM agents that uses checkpoints to distinguish failures occurring before and after a model’s capability is exposed, and applies controlled interventions to pinpoint upstream bottlenecks. The methodology is tested on four task families—delayed reuse of discovered information, reuse of observed state, recovery from failed strategies, and decision making after uncertain outcomes—revealing that many failures happen before the agent observes the state it later needs to reuse. Experiments with Gemini 2.5 Flash and Gemini 3.7 Flash show that targeted protocol‑disambiguation guidance can significantly alter state observation rates and that the primary source of failure can shift across model generations, underscoring the need for fine‑grained failure diagnostics rather than relying solely on overall task success.
By Wei Shao, Chongzhou Fang, Zuxiong Tan, Zequan Liang, Setareh Rafatirad, Avesta Sasan, Houman Homayoun
arXiv:2605. 12705v2 Announce Type: replace Abstract: How can we train models whose post-trained capabilities survive subsequent fine-tuning?
By Lawrence Feng, Gaurav R. Ghosal, Jacob Mitchell Springer, Ziqian Zhong, Aditi Raghunathan