NarrativeWorldBench evaluates 21 LLMs on nine narrative-structure metrics across horizons of 10 to 200 episodes, with cross-lingual support in Hindi, Tamil, Telugu, and Marathi. N-VSSM, a latent world model using Mamba-2, achieves plot-beat F1 of at least 0.84 across all horizons with 4x lower compute than closed-frontier models and outperforms Claude Opus 4.5 in long-arc consistency and controllability in a professional writer study.
Well-known brands dominate LLM recommendations by 100% when products are identical, but this advantage vanishes with a mere +0.1-star rating edge. Authority-style marketing claims, such as fabricated clinical evidence, break this dominance at a bias surplus of +0.17 rating points, with models responding differently. A social dilemma emerges in multi-brand competition, where collective optimization reduces individual payoff from +0.802 to +0.007 and eliminates recommendations for non-participating brands.
PARSE reduces prompt injection attack success from 25.4% to 15.6% on real enterprise documents across five professional domains, with statistically significant improvement (p=0.014) and 86.9% utility. It outperforms paraphrasing and uses provenance-aware sanitization to preserve factual content while routing most documents through a lightweight path.
AIPatient Arena evaluates large language models in end-to-end clinical consultations using EHR-grounded patient-specific knowledge graphs. It assesses LLMs across eight clinical competence dimensions, revealing strong performance in interview skills, ethics, and explanation clarity, but persistent weaknesses in handling ambiguity, information coverage, and diagnostic reasoning, with process failures like repetitive questioning and omitted history.
STATEWITNESS introduces an activation explainer that audits deception in reasoning LLMs by reading hidden states and generating natural-language answers or structured reports. It achieves a 0.916 mean AUROC, outperforming existing black-box monitors and activation probes by 11.6% and 25.0% respectively, and provides query-level, schema, and evidence-level traces for human inspection.
A new study identifies second-order bias in large language models—social bias in their judgments about biased content. Using entitlement epistemology, the research develops a reasoning task to assess whether LLMs accept or reject biased texts based on demographics, revealing implicit biases that vary by target group and evade safety guardrails. The work introduces two metrics to quantify these biases and calls for more theoretically grounded evaluation methods in NLP.
As enterprise agent tool catalogs scale from 10 to 110 agents, routing accuracy drops 16--23 percentage points on under-specified requests. An oracle analysis identifies retrieval and confusion gaps, with embedding-based shortlisting recovering +10--11pp F1. A human-annotated study of 1,435 utterances confirms real-world recovery of +10--17pp despite lower absolute performance.
This paper analyzes deep transformer expressiveness using bounded-depth grammars. It constructs transformers with positional attention where model depth scales linearly with grammar depth, and neuron count grows quadratically with production rules. The results support the linear representation hypothesis by showing these models can encode abstract grammatical states in low-dimensional, linearly separable subspaces.
NAR-MBR decoding improves speech recognition by maximizing expected utility from sampled outputs of non-autoregressive models. It achieves better performance than prior NAR methods and runs faster than autoregressive decoding across multiple corpora.
Concatenating LLM-generated features to graph neural networks systematically reduces accuracy on homophilous benchmarks, with PubMed accuracy dropping by -17.0 ± 0.3 pp. This degradation is linked to LLM-alone discriminability (Delta_sig), which correlates strongly with concatenation cost (r² = 0.38) and shows a power law relationship with feature dimension and node count (r² = 0.97), particularly in low-Delta_sig, low-node scenarios.
Pruned large language models often pass multiple-choice tests but fail to generate correct answers in open-ended responses. This 'benchmark illusion' shows that answers are not erased but demoted, reappearing only with advanced generation techniques like beam search or sampling. Standard benchmarks overstate the practical usability of compressed models, highlighting a critical evaluation blind spot.
OPD-Evolver introduces a slow-fast co-evolution framework that enables agents to select, act on, and reuse experience through on-policy self-distillation. It outperforms existing memory and training-based methods by up to 11.5% and 5.8% respectively, and demonstrates capability to challenge large-scale models like Qwen3.5-397B-A17B and Step-3.5-Flash.
A new framework uses prompt perturbation to identify and filter structurally inconsistent pairwise comparisons in large language model evaluations. By incorporating graph-level consistency checks before ranking aggregation, the method reduces cyclic preferences and improves the reliability of LLM rankings.
SkillMigrator learns reusable web skills by matching layout structures instead of specific element references. It stores each skill as a transferable interaction pattern (TIP) with a structural sketch, enabling efficient skill reuse across sites. Compared to state-of-the-art methods, it reduces average LLM-action counts by 8-10% on WebArena and Mind2Web at matched success rates.
MambaCount introduces a spatial sparse state space duality block to enable efficient text-guided open-vocabulary object counting. It addresses causal modeling limitations and high entropy in spatial token responses, achieving state-of-the-art results on FSC-147 with a test MAE of 12.23 while maintaining linear complexity.
SuCo introduces Minimal Sufficient CoT (MSC) as the shortest reasoning prefix adequate for correct answers. It employs a two-stage training framework—MSC-Aligned Fine-Tuning and Sufficiency-Aware Policy Optimization—to reduce reasoning length while maintaining or improving accuracy across math, code, and science tasks.
LLMs can detect cultural cues and recall cultural conventions, but often fail to adapt responses accordingly. Their responses remain biased toward their native culture unless explicitly prompted to apply cultural context sequentially.
EComAgentBench introduces a benchmark of 662 real Amazon tasks that scatter shopper requirements across query, profile, and clarification. Agents must uncover hidden intent, verify candidates with evidence, and commit to a product within 100 tool calls, with typed rubrics attributing failures to specific requirement sources. Evaluation shows even top models achieve only 57.1% accuracy, and rubric satisfaction drops when intent is hidden.
Current coding benchmarks were designed before agentic software engineering and fail to capture the complexity of real-world systems. They conflate model performance with the entire harness, ignore valid alternative solutions, and lack feedback signals at individual component levels, making iterative improvement difficult.
DIFE evaluates backdoored CLIP checkpoints across different deployment interfaces, revealing that native success does not guarantee safety in reuse. The framework shows text-side poisoning enables adversarial exposure in retrieval, reranking, and selection tasks, while visual-only use remains largely unaffected. BadTextTower is introduced to generate strong text-conditioned exposure without compromising visual performance.