The paper investigates how large language models balance instruction-following with pattern completion when the two objectives conflict. By creating dialogues where a user instruction to act in a target way T is opposed by assistant turns that demonstrate a competing pattern P, the authors measure instruction-following rates across 13 models and 16 instructions over up to 50 turns. Results show wide variability (1%–99%) in instruction adherence, with robustness influenced by instruction content, output format, and chain-of-thought reasoning, but overall instruction-following remains brittle under induction pressure.
By Carolina Camassa, Derek Shiller
arXiv:2608.28648v1 Announce Type: new
Abstract: We study how instruction-tuned LLMs arbitrate direct conflicts between system and user instructions. We introduce a benchmark of 41 paired constraints...
By Enrique Balp-Straffon, Chih-Hao Hsu, Rushiraj Gadhvi, Sunishchal Dev, Callum Stuart McDougall, Anusha Mujumdar
arXiv:2608. 11797v1 Announce Type: new Abstract: Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap.
By Chencheng Zhu
arXiv:2604. 04385v5 Announce Type: replace-cross Abstract: We localize the policy routing mechanism in alignment-trained language models.
By Gregory N. Frank
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.
By Nathan Labiosa, David Buff, Ena Nayak, Erica Donno
The study investigates the consistency and specificity of language model circuits across six tasks and five models, focusing on component-level (attention heads and MLP blocks) and neuron-level circuits. Component-level circuits are highly consistent and causally important but lack task specificity, as ablating a circuit for one task similarly harms performance on other tasks. Neuron-level circuits show higher task specificity but lower consistency, with overlap mainly between closely related tasks. The analysis of Llama‑3.2‑3B reveals that shared components are predominantly MLP blocks, while attention heads act as generic attention‑sink heads.
By Michael Li, Nishant Subramani
arXiv:2606. 07808v1 Announce Type: new Abstract: Reasoning language models deployed in agentic workflows must follow an instruction hierarchy: when instructions from different sources conflict, the model should obey the highest-privilege applicable instruction.
By Sanjay Kariyappa, G. Edward Suh
The paper investigates where the ‘refusal’ behavior of language models resides across different architectures. It finds that a single direction in the residual stream governs refusal in transformers, and that the same direction—after a rigid rotation—also governs refusal in state‑space models (SSMs). By aligning these directions and applying a detector‑triggered gate, the authors demonstrate that refusal can be effectively transferred across transformer, SSM, recurrent, and hybrid architectures, showing that safety tooling can be ported by re‑estimating the direction at each architecture’s write site rather than rebuilding it from scratch.
By Preethi Carmel Bosco, Gopalakrishnan Srinivasan
arXiv:2606. 05976v2 Announce Type: replace Abstract: Recent works show that LLM agents struggle to correct errors in their own reasoning traces, despite their ability to correct errors from external sources.
By Kuan-Yen Chen, Fang-Yi Su, Shih-Yen Lin, Bao Li, Jung-Hsien Chiang
arXiv:2605. 07284v2 Announce Type: replace Abstract: A late-layer change learned during post-training may work on the base model's earlier state, or it may depend on earlier computation learned with it.
By Yifan Zhou
arXiv:2606. 13705v1 Announce Type: cross Abstract: Yes.
By Aristotelis Lazaridis, Aman Sharma, Dylan Bates, Brian King, Vincent Lu, Jack FitzGerald
The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
By Zhenyan Lu, He Wang, Xiaohui Huang