LMEnt is a toolkit designed to study how language models acquire and represent world knowledge. It includes a richly annotated pretraining corpus, an improved entity-based retrieval method, and twelve pretrained models with numerous checkpoints. The suite enables controlled experiments linking entity mentions in training data to downstream knowledge performance, revealing how co-occurrence patterns influence learning and editing difficulty.
arXiv:2512. 04144v2 Announce Type: replace Abstract: Targeted interventions on language models, such as unlearning or model editing, aim to modify specific information, but their effects often propagate to related, unintended areas (e.
arXiv:2607. 07707v1 Announce Type: cross Abstract: Limited memory language models (LMLMs) externalize factual knowledge during pretraining to a knowledge base (KB), rather than memorizing it in their weights.
The paper investigates how language models acquire knowledge over the course of pretraining by using point-in-time models trained on web snapshots from different years. It introduces a temporal variability score (TVS) that measures how consistently a knowledge item is represented across these snapshots. The study finds that items with high TVS improve more rapidly later in training and outperform low‑TVS items in most benchmark settings, especially on knowledge‑intensive tasks.
The paper introduces FOVEATED, a plug‑and‑play framework that improves atomic‑fact recall in unstructured knowledge editing (UKE) for large language models. By randomly shifting Rotary Position Embedding (RoPE) positions during editing, FOVEATED creates focused views of each sentence, counteracting the context‑reliance problem where edited LLMs reproduce passages but fail to recall individual facts. Experiments show consistent gains across five editors, two LLM backbones, and three benchmarks.
The paper introduces a method for identifying the dominant knowledge source behind large language model (LLM) outputs, distinguishing between faithfulness violations (misuse of provided context) and factuality violations (errors in internal knowledge). A simple linear probe trained on hidden representations can reliably classify this source, and the authors present AttriWiki, a self‑supervised pipeline that generates labeled training data by prompting models to recall withheld entities or read them from context. Probes trained on AttriWiki achieve high Macro‑F1 scores across several models and datasets, generalize zero‑shot to a benchmark, and show that attribution mismatches can increase error rates by up to 70%. "whyItMatters":"The study demonstrates that knowing the source of an LLM’s answer is crucial for effective mitigation of hallucinations, as attribution mismatches significantly raise error rates."
arXiv:2607. 20433v1 Announce Type: cross Abstract: While language models remain frozen at their training state, the world evolves continuously.
The paper evaluates large pre‑trained language models (LLMs) for knowledge entity relation extraction (KERE) from academic articles, comparing multiple strategies such as closed‑source LLM CoT prompting, LoRA‑adapted open‑source LLMs, and various LLM‑enhanced small pre‑trained language models (SLMs). Two manually annotated datasets in the NLP domain—one for abstracts and one for full texts—were used, and the LoRA‑adapted T5 model achieved the highest macro F1 scores (0.7408 for abstracts and 0.7013 for full texts). Error analysis shows that LLM‑based strategies mitigate SLM errors differently across error types, and combining complementary strategies further improves KERE performance.
arXiv:2604. 00715v2 Announce Type: replace-cross Abstract: Retrieval-augmented generation (RAG) improves language model (LM) performance by providing relevant context at test time for knowledge-intensive situations.
The paper introduces “Rel-Bloom,” a multilingual diagnostic benchmark of 450 multiple-choice questions that tests language models’ ability to reason over relational facts supplied directly in prompts. It evaluates 43 models across six languages and two multilingual settings, revealing that accuracy drops sharply from Apply to Analyze levels, that Evaluate items remain a bottleneck even for high-performing models, and that difficulty rankings are consistent across languages. The study also finds that counterfactual contexts are generally harder than factual ones, and that removing entity identities can sometimes improve higher-level performance.
arXiv:2609.24238v1 Announce Type: new Abstract: We reproduce and stress-test the work of Yu et al. (2023), who characterize how language models (LMs) arbitrate between memorized knowledge and contrad...
arXiv:2608.30987v1 Announce Type: cross Abstract: Supervised fine-tuning (SFT) trains a base language model to imitate target responses, and these targets may require knowledge the base model has not...
The paper introduces ElephantBench, a closed‑book knowledge probe with 1,094 multi‑account factual questions generated via an auditable graph‑based pipeline that pulls documents from a low‑exposure web corpus and identifies naturally occurring disagreements. Across 32 large language models, even the best model only recovers both divergent accounts on 52.4% of questions, and most models recall one account while omitting the other, indicating persistent epistemic myopia. The study shows that scaling model size and inference‑time reasoning improves recall but does not eliminate incompleteness, and that exposure imbalance in the corpus biases models toward the dominant account.