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.
MODE-RAG proposes a multi-agent system using Variational Free Energy to dynamically gate interventions and reduce cross-modal hallucinations in retrieval-augmented generation. It integrates Monte Carlo Tree Search and logit perturbations to address causal fabrications and sycophancy, with dedicated agents ensuring factual verification and formatting stability. Evaluated via ModeVent, a subset of MultiVent, the system significantly improves robustness against logical fabrications.
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.
Large language models outperformed humans and supervised models in next speaker prediction using the AMI corpus, despite lacking audio-visual data and domain training. Multimodal LLMs surpassed text-based LLMs in addressee and turn-change detection but still fell short of human performance, highlighting challenges in utilizing raw audio-visual signals. Ablation studies show conversational context is crucial, especially for next speaker prediction, with both humans and LLMs struggling during frequent turn changes.
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.
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.
EnvRL introduces a framework that enhances agentic reinforcement learning by incorporating environment dynamics through state prediction and inverse dynamics objectives. It achieves significant gains in success rates on long-horizon benchmarks, improving Qwen-2.5-1.5B-Instruct performance from 72.8% to 77.4% on ALFWorld and from 56.8% to 67.0% on WebShop when trained with GRPO.
The LLM-as-Environment-Engineer framework uses LLMs to automatically redesign training environments in reinforcement learning by analyzing failure trajectories and contextual data. On the MAPF-FrozenLake testbed, it outperforms larger proprietary LLMs and fixed-environment baselines, with Qwen3-4B achieving the strongest aggregate performance. Analysis shows that failure evidence and preserved working configurations are key, and the current RL checkpoint performs better than the base model as an environment engineer.
SwiftTrans addresses runtime efficiency gaps in LLM-based code translation by introducing Multi-Perspective Exploration and Difference-Aware Selection. The framework extends CodeNet, F2SBench, and introduces SwiftBench to evaluate runtime performance, showing consistent improvements in both correctness and efficiency across benchmarks.
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.
A causal audit shows that text-only models match multimodal models in chest radiography accuracy. Across nine systems, a text-only model performs within 5.7 points of the best multimodal model, and a 119-billion-parameter model is indistinguishable from a 7-billion-parameter text-only baseline. Grounding audits, not accuracy, should determine clinical deployment.
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.