Classical and neural NLP: translation, question answering, tokenization and the evaluation of language understanding.
arXiv:2608.27971v1 Announce Type: new Abstract: Unmanned aerial vehicle (UAV) multimodal perception integrates visible (RGB), infrared (IR), synthetic aperture radar (SAR), and depth sensors for scen...
The paper explores using language models to restore missing text in damaged ancient manuscripts caused by physical gaps. It tests various scenarios, model architectures, and decoding strategies to handle tokenization mismatches and lacuna length awareness. Results show that while full automation is not yet possible, these tools can effectively aid paleographers, with performance varying by document section and missing text length.
The paper proposes the Effectiveness–Losslessness Framework to guide tokenization in domains beyond language, using predictive codelength as a criterion. It introduces two boundaries: the Fact–Token Boundary, where observable structure should be encoded into tokens, and the Token–State Boundary, where context‑dependent relations should remain for model state rather than being pre‑tokenized. Experiments on symbolic music show that making musical time explicit and applying tonal‑frame canonicalization improve predictive performance, while fixed pitch coordinates and reversible BPE can increase predictive code length, indicating that carrier compaction alone does not guarantee better predictions.
The paper introduces a new protocol, Mistake Detection Video Question Answering (MD‑VQA), to evaluate whether models can determine if a step in a video follows its description, covering both seen and unseen actions. It proposes a post‑training approach for video‑language models that uses a reward function to highlight discrepancies between instructions and video content. Experiments show this method surpasses zero‑shot, fine‑tuned, and other post‑training baselines, especially on unseen procedures, improving performance by up to 11.6% on EP‑VQA.
QUORUM is a budget‑aware routing framework that dynamically assigns each annotation instance to either human or large language model (LLM) annotators, using feature‑based signals to estimate difficulty and multiple annotations per instance. It combines annotations through agreement‑based rewards to improve reliability, without relying on model confidence or uncertainty estimates. Experiments on diverse closed‑ and open‑ended tasks in English and multilingual settings show that QUORUM can boost annotation quality by up to 34.4% while cutting costs by 8.8% compared to existing methods.
Entity-Memory graph retrieval preserves dialogue turns as verbatim memory nodes, links repeated mentions via shared entities, and connects adjacent memories with chronological edges. During retrieval, the system gates through entities, fuses semantics, and performs one‑hop chronological recovery before dense backfill, allowing it to keep neighboring memories that dense cosine ranking might miss. On 1,986 questions from ten LoCoMo conversations, this graph retrieval method increases official evidence recall at top‑k 25 from 79.7468 % to 84.4842 %, with the advantage extending from top‑k 5 to 50, though it does not improve overall final‑answer F1.
arXiv:2608.22849v2 Announce Type: replace Abstract: Full-length RNAs, particularly messenger RNAs, often exceed the context lengths used to pretrain existing RNA foundation models, limiting complete-...
The paper introduces the Speech-Unsupported Rejection Evaluation Challenge (SURE‑Challenge), a benchmark designed to test whether speech‑LLMs should accept or reject audio inputs before generating answers. Using LibriSpeech‑derived transcriptions paired with first‑word question answering, the authors evaluate various noise and silence conditions, and compare a simple energy‑plus‑Whisper‑score rule against a Qwen2‑Audio front‑end. On a 474‑row test set, the rule rejects 196 of 204 unsupported inputs while preserving accuracy on supported data, revealing a pre‑generation error mode that answer‑only scoring misses.
PRISM is an agentic retrieval framework that uses large language models in a structured loop to improve evidence gathering for multi‑hop question answering. It splits retrieval into three specialized agents—a Question Analyzer, a Selector focused on precision, and an Adder focused on recall—whose iterative interaction yields a compact yet comprehensive evidence set. Experiments on HotpotQA, 2WikiMultiHopQA, MuSiQue, and MultiHopRAG show that PRISM consistently outperforms strong baselines by achieving higher retrieval accuracy and filtering out distracting content.
arXiv:2608.28329v1 Announce Type: cross Abstract: Medical question answering (QA) systems have become crucial tools for providing reliable health information. But they remain very unexplored for low-...
ContextPilot is a proactive context‑management framework designed to improve long‑horizon agentic reasoning with large language models. It expands the toolset to include planning, long‑term memory, and soft context offloading, and introduces a reinforcement‑learning strategy that focuses on critical editing decisions and assigns action‑level advantages. Experiments on long‑context QA and deep search tasks demonstrate that ContextPilot achieves stronger performance with a more compact working context, outperforming existing baselines across various base models and benchmarks.
The paper introduces Synth-JDoc, a synthetic Japanese document image dataset created by rendering text with HTML and CSS to produce multi‑column layouts that include both vertical and horizontal writing styles. Images generated by text‑to‑image models are embedded to enhance visual realism, and noise and degradation filters are applied to improve robustness. Experiments show that fine‑tuning Large Vision Language Models on Synth‑JDoc yields superior performance on reading vertically written Japanese text compared to prior synthetic datasets.
The paper introduces a distance‑guided decoding framework that guarantees context‑free grammar compliance for large language models. By precomputing bounded pushdown summaries with reachability labels and upper‑bound distances to acceptance, the decoder performs horizon‑aware pruning and beam search during generation. Experiments on JSON, SQL, and Linear Temporal Logic demonstrate consistent syntactic validity and improved completion quality compared to existing baselines.
Doc‑CoB introduces a Chain‑of‑Boxes framework that enhances document understanding by progressively focusing on query‑relevant layout regions while preserving global context. It selects key layout boxes and then applies visual prompting for deeper analysis, supported by two new reasoning tasks and an automatic pipeline that generates 249k training samples with intermediate visual supervision. Experiments across seven benchmarks and four popular models demonstrate significant performance gains, underscoring the method’s effectiveness and broad applicability.
The paper examines two test‑time scaling methods for large language models in machine translation: sequential sampling, where later attempts build on earlier ones, and parallel sampling, such as independent i.i.d. sampling with reranking. Sequential sampling shows a higher performance ceiling, offering a more diverse and effective set of translations, especially with limited sampling budgets. Human analysis reveals that while sequential sampling improves fluency and naturalness, it can reduce accuracy when the inference budget is large, and the authors attribute this effect to the model’s access to a larger target‑side context.
XHotpotQA is a new benchmark for cross‑lingual knowledge composition in multi‑hop question answering. It presents each instance as an evidence‑dependency graph with explicit language assignments for the question, bridge evidence, answer‑bearing evidence, and distractors, and includes 15,661 training and 7,405 validation examples with sentence‑level support supervision. The dataset reveals significant performance drops when evidence spans language boundaries, providing a diagnostic tool for systems that must integrate evidence across languages.
The paper shows that post‑training quantization can introduce backdoors in large language models that are not detected by source‑precision checks. By formalizing the validation‑deployment gap with Quantization Behavioral Equivalence Classes (QBECs), the authors demonstrate that models can pass full‑precision tests yet exhibit malicious behavior after INT8 or 4‑bit compression. Experiments on machine translation and political stance classification reveal significant corruption and ideological shifts, and cross‑quantizer analysis indicates that attack persistence depends on the quantization scheme and architecture rather than just bit‑width.
The UIC-AIHealth4All system was presented for the ArchEHR-QA 2026 shared task on grounded question answering from electronic health records. It participated in evidence identification, answer generation, and answer‑evidence alignment, using an answer‑first pipeline that generates candidate answers with cited note sentences before classifying the full evidence set. The system ranked third in evidence identification, ninth in answer generation, and fifth in answer‑evidence alignment, and a linguistic analysis showed its outputs were harder to read than clinician‑authored references, highlighting the need for readability optimization in clinical NLP.
The paper introduces CodeRQ-Bench, the first benchmark for assessing large language model reasoning quality across coding tasks such as generation, summarization, and classification. It analyzes over a thousand mismatches from existing evaluators, identifies recurring limitations, and derives design insights that lead to a new two‑stage evaluator, VERA. Experiments show VERA outperforms strong baselines, improving AUCROC by up to 0.26 and AUPRC by up to 0.21 on four datasets.
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.