The paper shows that chain‑of‑thought (CoT) instructions can distort multiple‑choice vision‑language model evaluation when a scorer appends a reasoning cue but reads answer‑label logits before the model generates any rationale. This CoT‑prefix scoring causes significant drops in accuracy (e.g., Qwen2.5‑VL‑7B falls from 80.76% to 45.48% on ScienceQA) and leads most predictions to choose the first option. Analysis reveals that while answer information remains linearly accessible in late layers, the immediate readout is misled by probability mass shifting toward continuation tokens, and the issue varies across datasets and models.
arXiv:2606. 27359v1 Announce Type: cross Abstract: Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level.
By Johannes Zenn, Jonas Geiping
The paper shows that chain‑of‑thought (CoT) instructions can distort evaluation of vision‑language models (VLMs) when a scorer reads answer‑label logits before the model generates a rationale. On ScienceQA, Qwen2.5‑VL‑7B’s accuracy falls from 80.76% to 45.48% under this CoT‑prefix scoring, and most predictions incorrectly pick the first option. Linear probes and free generation recover most of the lost accuracy, indicating that the answer information remains in the hidden states but is missed by the early readout. The authors explain the mismatch with vocabulary and layer diagnostics, noting that probability mass shifts toward continuation tokens while answer information stays linearly accessible in later layers. The effect varies across datasets and models, but the study demonstrates that CoT‑prefix scoring can misrepresent model knowledge unless the requested and scored outputs are aligned.
By Zeyan Li, Siyuan Qiu, Jianfeng Xu
Language models increasingly use skills that provide reusable instructions and procedures for task execution. After a relevant skill has been selected, the model may need to apply its governing proced...
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
arXiv:2606. 28337v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) systems are often evaluated using final answer accuracy, even though their failures can originate from preprocessing, retrieval, context packing, or generation.
By Bharath Simha Reddy Muthyam
arXiv:2609.07309v1 Announce Type: cross
Abstract: Robustness evaluation of large language models (LLMs) remains a critical challenge, particularly in assessing their sensitivity to perturbations in i...
By Vamsi Krishna Kodavali, Rituraj Singh
The paper demonstrates that multiple‑choice visual question answering (MC‑VQA) benchmarks are unreliable because model performance is highly sensitive to semantically neutral prompt formatting choices—such as option ID sets, delimiters, and separators—despite protocols that mitigate option‑order effects. Across seven multimodal large language models and five datasets, the authors observed frequent rank reversals when systematically varying 48 equivalent prompt formats, attributing the instability to tokenizer‑induced token fusion or removal and to how option ID sets influence attention patterns. Consequently, MC‑VQA rankings correlate weakly with open‑ended evaluation, revealing that MC‑VQA reflects option‑selection dynamics as well as multimodal reasoning.
By Fabio Rosenthal, Sebastian Schmidt, Thorsten Graf, Thorsten Bagdonat, Stephan G\"unnemann, Leo Schwinn
arXiv:2610.35826v1 Announce Type: new
Abstract: Selection bias in large language models (LLMs) can persist despite instructions to ignore presentation order or prompts that elicit reasoning. Recent e...
By Mingyang Song, Mao Zheng, Zhiyuan Ma, Xinpeng Zhou
arXiv:2610.19185v1 Announce Type: new
Abstract: Evaluation of large language models (LLMs) remains largely pre-scientific: it involves many design decisions, with little guidance on how to make them....
By Shenbin Qian, David Samuel, Jind\v{r}ich Helcl, Andrey Kutuzov, Stephan Oepen
arXiv:2603.13911v2 Announce Type: replace
Abstract: A language model can give the wrong answer even when the correct answer is decodable from its intermediate states. To study this gap between decoda...
By Valeria Ruscio, Keiran Thompson
arXiv:2607. 09438v1 Announce Type: cross Abstract: Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear.
By Spiros Baxevanakis, Peng-Jian Yang