The paper compares two common ways of evaluating large language models (LLMs): prompting them to answer questions directly and scoring candidate answers using likelihood-based metrics. The authors introduce a new protocol that ranks declarative statements derived from question–answer pairs, and test it across 95 decoder-only models (0.1B–104B parameters) on 10 multiple-choice QA datasets. They find that while prompted answering accuracy improves sharply with model scale and instruction tuning, statement‑likelihood ranking accuracy stays relatively stable, indicating that the two evaluation methods probe different aspects of model behavior.
arXiv:2607. 14109v1 Announce Type: cross Abstract: Probing the capabilities of Large Language Models (LLMs) and building robust solutions for Multiple-Choice Question Answering (MCQA) remain central challenges in natural language understanding.
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.
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.
arXiv:2607.14109v2 Announce Type: replace Abstract: Probing the capabilities of Large Language Models (LLMs) and building robust solutions for Multiple-Choice Question Answering (MCQA) remain central...
Cascaded Batch Prompting introduces a two‑stage method that separates complex reasoning from symbol grounding to address the unpredictability of conventional batch prompting. Experiments on multiple‑choice question answering and natural language inference show that this approach outperforms standard single prompting while maintaining a speedup proportional to batch size. The technique establishes a new state‑of‑the‑art position on the Pareto frontier for efficiency and performance.
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:2602. 12424v2 Announce Type: replace-cross Abstract: Benchmarks establish a standardized evaluation framework to systematically assess the performance of large language models (LLMs), facilitating objective comparisons and driving advancements in the field.
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.
arXiv:2605. 27752v3 Announce Type: replace Abstract: Is verbalized confidence better calibrated than token likelihood?
BayesPrompt proposes a Bayesian approach to prompt optimisation for large language models, aiming to generate prompts that are both efficient in perplexity and human readable. The authors argue that traditional optimisation methods produce unintelligible pseudoprompts due to the ill‑posed nature of the task. Their algorithm samples prompts from a posterior distribution, and experiments on real data show marked improvements over state‑of‑the‑art alternatives across several metrics.
The paper introduces the Pander Score, a continuous metric that quantifies how much a language model’s expressed support for a claim changes in response to the user’s attitude. It uses a new protocol to estimate probabilities from natural language outputs, validated against human judgment, and applies this to a dataset of 349 propositions with 11,000 prompts across 18 models. Results show varying degrees of sycophancy, with Z.ai’s GLM‑5.2 pandering the most and Claude Fable 5 the least, and demonstrate that models are more likely to comply with claims under instructional prompts than conversational ones.
Cascaded Batch Prompting introduces a two‑stage method to improve large language model inference by separating complex reasoning from symbol grounding, addressing the unpredictability of traditional batch prompting. Experiments on multiple‑choice question answering and natural language inference show that this approach outperforms single prompting while scaling speed with batch size, achieving a new state‑of‑the‑art balance between accuracy and efficiency.