The study investigates how Google’s Gemma 4‑e4b language model resolves conflicts between two documents. Using a counterbalanced design, researchers found that the semantic framing of a source (e.g., labeling it as an official guideline) dominates over the order in which documents appear. While the model shows a primacy bias toward the first document, this bias varies widely with wording and is amplified only when the documents are structurally identical.
arXiv:2605. 26795v2 Announce Type: replace Abstract: Chain-of-thought (CoT) prompting enhances large language model performance, yet what drives these gains remains unclear.
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...
arXiv:2606. 07555v1 Announce Type: cross Abstract: Glossaries, technical specifications, and system prompts routinely ask language models to use familiar words in unfamiliar ways.
arXiv:2601. 04098v2 Announce Type: replace-cross Abstract: Transformer language models systematically prefer tokens at specific input positions regardless of semantic relevance---a phenomenon known as positional bias.
The paper introduces the concept of summarization bias in large language models (LLMs), describing a systematic tendency for LLMs to represent narrative meaning as an abstract summary label rather than the reconstructable inferential structure that produces it. It frames this bias within the Bulut Doctrine’s told‑shown axis, arguing that LLMs fail in a specific direction: they default to told‑mode explicitness in generative tasks and reward told‑mode explicitness while under‑detecting shown‑mode suppression in evaluative tasks. The authors outline two regimes of bias, present preliminary evidence, and pre‑register a test protocol to validate or abandon the construct.
arXiv:2608. 12387v1 Announce Type: cross Abstract: Positional biases such as recency and primacy effects have been documented in large language models (LLMs), yet the underlying mechanism by which these models make their evaluations remains poorly understood.
arXiv:2608. 04021v1 Announce Type: cross Abstract: Cloze-style probes that vary how often a target token appears implicitly assume that more copies of a target affect prediction the same way regardless of where the readout slot sits.
The study examines how different editorial framings in prompts influence large language models’ statistical analysis reports. Using a 4×4 factorial design, researchers found that certain framings—particularly brutally critical prompts on genuine effects and significance-seeking prompts on underpowered nulls—led to factual misrepresentations. Tone shifts were more widespread, with critical framing inducing defensive language across all data patterns, while a confound in the data largely prevented both factual and tonal distortions.
arXiv:2607. 21692v1 Announce Type: new Abstract: Sparse attention reduces the cost of long contexts by allowing each query to read only selected parts of the input.
arXiv:2609. 11291v1 Announce Type: new Abstract: We post-train Qwen3.
The paper investigates how language models decide between contextual information and their internal memory when the two conflict. By estimating "authority directions" from agreement prompts and swapping these directions between matched prompts, the authors show that such interventions can reproduce 30–68% of the shift in source choice across Qwen, Llama, and OLMo models. Cross‑task experiments reveal that authority directions learned on one task transfer only modestly (≈9%) to another, indicating that authority computations are largely task‑specific.
The paper introduces a bias depth score to differentiate between stable model preferences (Deep biases) and prompt‑dependent responses (Shallow biases) in large language models. By analyzing 4,442 opinion prompts across four models, it finds that only about a quarter of concentrated preferences persist after scenario reframing, indicating that most are shallow. The study shows Deep biases are more often inherited from pretraining and harder to remove through fine‑tuning or prompt‑based debiasing, highlighting the need to distinguish learned biases from prompt artifacts.