The paper investigates the reliability of attention‑head ablation as a causal inference tool in language models. Using GPT‑2 small, the authors find that a natural post‑projection zeroing method is almost uncorrelated with a corrected pre‑projection ablation and yields a completely different set of top‑5 important heads. They also show that binary accuracy can mask effects near performance floors or ceilings, whereas gold‑token log‑probability provides a graded signal. By employing a discovery/held‑out split and 1,000 matched random‑head and layer‑matched‑head controls, the corrected per‑head effect ranking remains highly stable (Spearman ρ = 0.974) and the top‑5 heads significantly outperform both control distributions (Monte Carlo p = 0.001). However, evidence for task specificity is weak on GPT‑2, and replication on DistilGPT‑2 confirms the intervention‑semantic and matched‑control findings.
By Juli Huang
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.
By Qiji Zhou, Junyi Hu, Peiyu Wang, Yue Zhang
The paper investigates how large language models (LLMs) handle affective state reasoning, a type of implicit conversational state that persists across topic shifts. It introduces Self‑targeted Internal Attention Boosting (SIAB), a method that identifies and amplifies sparse affective attention heads using causal localization and the model’s own behavior, without relying on external labels. Experiments on ECC and GoEmotions datasets across multiple LLMs and prompting settings show that SIAB improves macro‑F1 scores by up to +10.3 points on ECC and +5.5 on GoEmotions, and further gains are achieved when combined with external supervision.
By Wenqiu Tang, Zhen Wan, Yihui Wang, Takahiro Komamizu, Ichiro Ide
arXiv:2608. 14509v1 Announce Type: new Abstract: Systems that ask a language model to reach a conclusion from many sources usually concatenate them into one prompt.
By Zhelun Wu
arXiv:2506.19549v2 Announce Type: replace-cross
Abstract: Identifying which tokens and activations truly influence model predictions is critical for both the efficiency and interpretability of large...
By Debabrata Mahapatra, Shubham Agarwal, Apoorv Saxena, Subrata Mitra
The study investigates how the position of evidence documents relative to a prompt affects long‑context question answering across different sequence mixers—Transformer, state‑space, and hybrid Mamba models. Results show that Transformers exhibit a beginning‑position advantage that state‑space models lack, while both families tend to favor evidence near the end. Additional experiments reveal that prompt design, corpus choice, and model capability influence performance, and that position bias remains decodable across architectures even when response curves differ.
By Aman Behera, Namit Solanki, Mehul Anand
Faithfully understanding the internal process that occurs as models form their chain of thought is an important problem with the rise in model capabilities. However, this remains challenging due to th...
The paper introduces Causal Attribution Pruning (CAP), a training‑free technique that identifies critical attention heads in large language models by measuring their causal impact on reasoning tasks. CAP estimates performance loss when each head is masked on a small calibration set, then translates these causal scores into weight‑level importance for pruning. Compared to magnitude‑only or activation‑based methods, CAP achieves up to 61% relative accuracy gains over Wanda on ARC‑Challenge at 20% sparsity and improves performance on GSM8K, StrategyQA, and ARC‑Challenge across Llama‑3‑8B‑Instruct and Mistral‑7B‑Instruct at 10%, 20%, and 50% sparsity.
By Amogh Sheth, Andrew Lin, Biruk Assefa, Yi Wen Huang, Yuhao Ge
The paper introduces a method for analyzing how interventions in a prompt affect a language model’s reasoning by using behavioral counterfactuals. It trains a module to edit the residual stream so that the next-token distributions for a fixed response match between a source and a counterfactual prompt, outperforming standard activation reconstruction. The resulting attribution maps and visualizations help identify causally important locations for counterfactual recovery and guide hypothesis testing with activation patching, as demonstrated in a case study on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B.
By Shawn Im, James Oldfield, Adel Bibi, Sharon Li
arXiv:2609.37616v1 Announce Type: new
Abstract: Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on th...
By Abhinav Rajeev Kumar (Lossfunk), Paras Chopra (Lossfunk)
arXiv:2607. 01002v1 Announce Type: cross Abstract: In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them.
By Aryo Pradipta Gema, Beatrice Alex, Pasquale Minervini
The paper introduces ICE (Intervention-Consistent Explanation), a framework that evaluates the faithfulness of large language model explanations by comparing them to random baselines of equal size across multiple intervention operators. It demonstrates that faithfulness varies with the chosen operator, with significant differences observed when switching between deletion and retrieval infill operators. The study evaluates seven LLMs on four tasks, revealing that operator changes can cross the positive-evidence threshold in 18% of configurations and that random baselines uncover anti-faithfulness in nearly a third of English deletion setups, findings that also hold across six non‑English languages and two attribution methods.
By Abhinaba Basu, Pavan Chakraborty