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:2609.01048v1 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 investigates how a probe can decode in‑context bindings on model errors and how probe‑guided steering can repair them. It tracks probe accuracy, model output, and steering response across public pretraining and post‑training checkpoints, noting that probe accuracy improves during Pythia pretraining and that steering benefits grow with model size. The study also shows that decoders trained on final state or candidate logits do not outperform each other on late‑checkpoint errors, and presents an information‑theoretic counterexample explaining why decodability on errors alone cannot prove discarded output information.
By Manas Venkata Sai Ravulapalli, Samrath Singh Chadha
The paper introduces Counteraction-Aware Multi-Teacher On-Policy Distillation (CaMOPD), a method designed to recover general capabilities in large language models while preserving domain-specific behavior. CaMOPD tackles two failure modes of standard Multi-Teacher On-Policy Distillation—conflicting recovery and preservation gradients, and weak correction signals—by using decoupled alternating training and selecting samples with large teacher‑student log‑probability gaps. Experiments on role‑play dialogue and medical reasoning QA show that CaMOPD outperforms baselines in general capability recovery while maintaining domain specialization, and gradient coherence analyses confirm more coherent correction signals.
By Tianlei Chen, Jiao Ou, Ziyuan Liu, Ruiming Tang, Jian Liang, Han Li
The paper investigates how the upper spectral tails of weight matrices in decoder‑only transformer language models influence reasoning behavior. By performing controlled interventions on the query–key product and comparing them to factor‑level surgeries, the authors find that edits targeting the spectral tail more strongly affect model performance across multiple checkpoints and reasoning benchmarks. The study also explores how inverse participation ratios predict accuracy transitions and shows that tail‑aware low‑rank adaptations converge faster than standard methods.
By Ibne Farabi Shihab, Sanjida Akhter, Md Najmus Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Anuj Sharma
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 investigates Contrastive Activation Addition (CAA), a method that steers language‑model behavior by adding activation vectors derived from contrasting examples. By framing CAA as a behavior‑capability frontier, the authors select interventions that maximize behavioral gain while limiting capability degradation to specified budgets (1%, 3%, 5%). Across four instruction‑tuned models and seven behaviors, they find that validated behavioral effects generalize well, and that joint behavior‑capability survival improves significantly with higher budgets, though additional gains plateau. Control experiments confirm that the effects rely on learned behavior directions rather than generic perturbations, and representational geometry only modestly predicts cross‑behavior transfer. Secondary tests reveal that even interventions meeting protected benchmarks can still degrade untested capabilities or transfer unevenly to free‑form generation, underscoring that activation steering is a constrained, model‑dependent intervention rather than a universal control knob.
By Deepanshu Raj
arXiv:2608. 14717v1 Announce Type: cross Abstract: A query-relation deletion can improve the edited slot while reducing the utility of the prediction set that contains it.
By Ze Zhang, Yang Zhang
The study investigates how layer‑wise intervention responses in language models change over the course of pretraining, using single‑block identity bypass across multiple checkpoints and model‑domain combinations. It finds that while depth ordering of responses persists, their magnitudes shift, with nearby checkpoints showing stronger rank correspondence than distant ones and large changes occurring at positions that recur across samples and transfer across evaluation domains. Controlled experiments reveal that these longitudinal changes cannot be explained by a single downstream sensitivity and depend on perturbation strength and direction, indicating that layer sensitivity is structured but dynamic.
By Shengye Tao, Yinzhu Cheng, Haihua Xie
arXiv:2607. 16821v1 Announce Type: cross Abstract: Task arithmetic, sequential fine-tuning, activation steering, and first-order random search all operate through relatively small perturbations around an already trained checkpoint, and they rely on different local approximations: individual perturbations should be first-order predictable, task updates should compose with controlled interference, useful tangent structure should be stable and possible to estimate, and weight edits should have counterparts in representation space.
By Irina Piontkovskaia, Sergey Nikolenko
arXiv:2607. 18553v1 Announce Type: cross Abstract: Can a language model read the quality of ongoing computation, and can an external intervention turn that readout into better outcomes?
By Jan Kirin
The paper introduces Amplitude Gating (AG), an inference-time technique that selectively masks feed‑forward network (FFN) activation coordinates while keeping pretrained weights unchanged. AG employs six SwiGLU sites to separate shared and gate‑only inputs, expands gate/value channels, and uses magnitude‑ranked entropy masks or a Newton‑Schulz windowed ranking variant to provide alternative forward paths without retraining. Experiments on three Qwen checkpoints and eight text/tool endpoints show modest improvements in function‑call and ToolBench endpoints, though statistical significance is not established and further experiments are required.
By Sheng Xu, Zhen Chen, Junhua Wang, Boyuan Huang, Ke Jia, Jiadun Zhu, Yiming Xu