MCP-Persona is a new benchmark exposing significant struggles of state-of-the-art LLM agents with personalized tool use across social and collaboration applications.
naacl-long.331/
51 Pith papers cite this work. Polarity classification is still indexing.
51
Pith papers citing it
citation-role summary
background 3
citation-polarity summary
years
2026 51roles
background 3polarities
background 3representative citing papers
DeepTool scales interleaved deliberation in tool-integrated reasoning via a synthesis pipeline for trajectories and GRPO-based process-supervised RL with an action-centric reward, reporting large gains on math benchmarks.
ClaimRAG-LAW is a French-English legal RAG benchmark with claim-level granularity for experts and non-experts that reveals limitations in current retrieval and generation performance.
SCICONVBENCH is a new benchmark evaluating LLMs on multi-turn disambiguation and inconsistency resolution for task formulation in computational science, with frontier models reaching only 52.7% success on fluid mechanics disambiguation cases.
Prompt-boundary directional alignment enables geometry-guided search that cuts trials to 95% best utility by 39.8% on average, while concept granularity predicts remaining difficulty via directional heterogeneity.
SxS Interleaved Reasoning learns when to disclose partial reasoning during generation and improves accuracy versus content-latency trade-offs on math and science benchmarks.
SKPO improves outcome-based RL for reasoning by adding skip connections that let models bypass flawed early reasoning while preserving access to the original problem, yielding 3.91-6.17% relative gains and higher-quality intermediate steps.
MuRGAt benchmark reveals that strong multimodal models frequently hallucinate citations in complex reasoning tasks despite correct answers, exposing a gap between internal reasoning and verifiable attribution.
xMemory builds revisable hierarchical agent memory by segmenting histories, decoupling into components, and aggregating via sparsity-semantic objective, yielding better answer quality and lower token use than flat RAG on LoCoMo and PerLTQA.
NeuNeu, a neural network trained on HuggingFace checkpoints, predicts language model accuracy on 66 downstream tasks at 1.99% MAE by extrapolating trajectories, outperforming logistic scaling laws by 44% and generalizing zero-shot to new models and tasks.
Heading-anchored steering vectors exert bidirectional causal control over tool-invocation behavior across five LLMs, but geometric analysis reveals diffuse, non-linear structure inconsistent with parametric concept encoding.
MBR decoding is reformulated via a noisy-channel decomposition into four weighted probabilistic terms, revealing that channel importance is metric-specific and task-agnostic, and that reweighting can improve performance.
A Judge-Aware Gated Multi-Task Learning architecture with outcome taxonomy supervision achieves SOTA accuracy on 13,937 UK Employment Tribunal decisions using an order of magnitude fewer parameters than generative SFT baselines on a 26B model.
A token-efficient method to curate high-quality reasoning SFT data using early loss patterns from perturbed checkpoints outperforms baselines on medical and math datasets.
SOLAR aligns soft-token probability mixtures across languages in embedding space during SFT and raises multilingual reasoning accuracy by up to 17.7 points over the base model.
AttriCoT is a black-box algorithm that attributes causal importance to units in a specific CoT trace via a structural causal model estimated with linear forward passes.
TLVS mitigates hallucinations in LVLMs via token-level extraction and visual-sensitivity-adaptive steering applied only at critical decoding steps.
Unlearning in multilingual LLMs suppresses rather than erases knowledge in later layers, with transfer varying by language similarity and reversible via inference-time steering.
IC-VCO places contrastive images in one context for a consistent DPO-style objective, adds Visual Contrast Distillation, and uses semantic perturbation for hard negatives, reporting best results on five benchmarks.
LearnWeak specializes small CUAs via weakness detection by a reference agent, targeted task synthesis, and error-aware training, delivering 11+ point gains on OSWorld.
CIE-Scorer detects unfaithful CoT by tracing compact sentence-level circuits, building internal-external reasoning graphs, and scoring their discrepancy with Fused Gromov-Wasserstein distance, reporting SOTA results on FaithCoT-Bench with reduced circuit cost.
PUMA detects reasoning-level semantic redundancy to enable early exit in chains of thought, achieving 26.2% average token reduction across five LRMs and five benchmarks while preserving accuracy and CoT quality.
LLMs exhibit Pseudo-Deliberation where explicit reasoning fails to align stated values with generated actions, measured via the new VALDI framework across 4,941 scenarios in five domains.
VLMs suffer from a perceptual bandwidth bottleneck; the paper formalizes active visual reasoning as sequential Bayesian optimal experimental design, derives a coverage-resolution proxy objective, and introduces the training-free FOVEA method that yields gains on high-resolution benchmarks.
citing papers explorer
-
MCP-Persona: Benchmarking LLM Agents on Real-World Personal Applications via Environment Simulation
MCP-Persona is a new benchmark exposing significant struggles of state-of-the-art LLM agents with personalized tool use across social and collaboration applications.
-
DeepTool: Scaling Interleaved Deliberation in Tool-Integrated Reasoning via Process-Supervised Reinforcement Learning
DeepTool scales interleaved deliberation in tool-integrated reasoning via a synthesis pipeline for trajectories and GRPO-based process-supervised RL with an action-centric reward, reporting large gains on math benchmarks.
-
Fine-grained Claim-level RAG Benchmark for Law
ClaimRAG-LAW is a French-English legal RAG benchmark with claim-level granularity for experts and non-experts that reveals limitations in current retrieval and generation performance.
-
SCICONVBENCH: Benchmarking LLMs on Multi-Turn Clarification for Task Formulation in Computational Science
SCICONVBENCH is a new benchmark evaluating LLMs on multi-turn disambiguation and inconsistency resolution for task formulation in computational science, with frontier models reaching only 52.7% success on fluid mechanics disambiguation cases.
-
When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search
Prompt-boundary directional alignment enables geometry-guided search that cuts trials to 95% best utility by 39.8% on average, while concept granularity predicts remaining difficulty via directional heterogeneity.
-
When to Think, When to Speak: Learning Disclosure Policies for LLM Reasoning
SxS Interleaved Reasoning learns when to disclose partial reasoning during generation and improves accuracy versus content-latency trade-offs on math and science benchmarks.
-
Skip-Connected Policy Optimization for Implicit Advantage
SKPO improves outcome-based RL for reasoning by adding skip connections that let models bypass flawed early reasoning while preserving access to the original problem, yielding 3.91-6.17% relative gains and higher-quality intermediate steps.
-
Multimodal Fact-Level Attribution for Verifiable Reasoning
MuRGAt benchmark reveals that strong multimodal models frequently hallucinate citations in complex reasoning tasks despite correct answers, exposing a gap between internal reasoning and verifiable attribution.
-
Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
xMemory builds revisable hierarchical agent memory by segmenting histories, decoupling into components, and aggregating via sparsity-semantic objective, yielding better answer quality and lower token use than flat RAG on LoCoMo and PerLTQA.
-
Neural Neural Scaling Laws
NeuNeu, a neural network trained on HuggingFace checkpoints, predicts language model accuracy on 66 downstream tasks at 1.99% MAE by extrapolating trajectories, outperforming logistic scaling laws by 44% and generalizing zero-shot to new models and tasks.
-
Controlling Tool Use with Heading-Specific Activation Steering
Heading-anchored steering vectors exert bidirectional causal control over tool-invocation behavior across five LLMs, but geometric analysis reveals diffuse, non-linear structure inconsistent with parametric concept encoding.
-
Noisy-Channel Minimum Bayes Risk Decoding
MBR decoding is reformulated via a noisy-channel decomposition into four weighted probabilistic terms, revealing that channel importance is metric-specific and task-agnostic, and that reweighting can improve performance.
-
Towards Explainable Adjudicative Variance: Quantifying Judicial Discretion via Gated Multi-Task Learning
A Judge-Aware Gated Multi-Task Learning architecture with outcome taxonomy supervision achieves SOTA accuracy on 13,937 UK Employment Tribunal decisions using an order of magnitude fewer parameters than generative SFT baselines on a 26B model.
-
Reasoning Quality Emerges Early: Data Curation for Reasoning Models
A token-efficient method to curate high-quality reasoning SFT data using early loss patterns from perturbed checkpoints outperforms baselines on medical and math datasets.
-
Soft Token Alignment for Cross-Lingual Reasoning
SOLAR aligns soft-token probability mixtures across languages in embedding space during SFT and raises multilingual reasoning accuracy by up to 17.7 points over the base model.
-
Local Causal Attribution of Chain-of-Thought Reasoning
AttriCoT is a black-box algorithm that attributes causal importance to units in a specific CoT trace via a structural causal model estimated with linear forward passes.
-
Steer Where It Matters: Token-Level Visual-Sensitivity Steering for LVLMs Hallucination Mitigation
TLVS mitigates hallucinations in LVLMs via token-level extraction and visual-sensitivity-adaptive steering applied only at critical decoding steps.
-
Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility
Unlearning in multilingual LLMs suppresses rather than erases knowledge in later layers, with transfer varying by language similarity and reversible via inference-time steering.
-
Learning from Fine-Grained Visual Discrepancies: Mitigating Multimodal Hallucinations via In-Context Visual Contrastive Optimization
IC-VCO places contrastive images in one context for a consistent DPO-style objective, adds Visual Contrast Distillation, and uses semantic perturbation for hard negatives, reporting best results on five benchmarks.
-
Learn from Weaknesses: Automated Domain Specialization for Small Computer-Use Agents
LearnWeak specializes small CUAs via weakness detection by a reference agent, targeted task synthesis, and error-aware training, delivering 11+ point gains on OSWorld.
-
Detecting Unfaithful Chain-of-Thought via Circuit-Guided Internal-External Discrepancy
CIE-Scorer detects unfaithful CoT by tracing compact sentence-level circuits, building internal-external reasoning graphs, and scoring their discrepancy with Fused Gromov-Wasserstein distance, reporting SOTA results on FaithCoT-Bench with reduced circuit cost.
-
Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models
PUMA detects reasoning-level semantic redundancy to enable early exit in chains of thought, achieving 26.2% average token reduction across five LRMs and five benchmarks while preserving accuracy and CoT quality.
-
Pseudo-Deliberation in Language Models: When Reasoning Fails to Align Values and Actions
LLMs exhibit Pseudo-Deliberation where explicit reasoning fails to align stated values with generated actions, measured via the new VALDI framework across 4,941 scenarios in five domains.
-
The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design
VLMs suffer from a perceptual bandwidth bottleneck; the paper formalizes active visual reasoning as sequential Bayesian optimal experimental design, derives a coverage-resolution proxy objective, and introduces the training-free FOVEA method that yields gains on high-resolution benchmarks.
-
Segment-Aligned Policy Optimization for Multi-Modal Reasoning
SAPO introduces segment-level policy optimization using a step-wise MDP abstraction to better align RL updates with reasoning structure in multi-modal LLM tasks.
-
HeadRouter: Dynamic Head-Weight Routing for Task-Adaptive Audio Token Pruning in Large Audio Language Models
HeadRouter prunes audio tokens more effectively by dynamically routing based on per-head importance for semantic versus acoustic tasks, exceeding baseline performance at 70% token retention on Qwen2.5-Omni models.
-
Shared Lexical Task Representations Explain Behavioral Variability In LLMs
Attention heads that decode to task-descriptive words in vocabulary space are shared across prompting styles, and their activation strength causally explains behavioral variance in LLMs.
-
Process Supervision via Verbal Critique Improves Reasoning in Large Language Models
Verbal Process Supervision uses structured critiques from stronger models in an iterative loop to improve LLM reasoning, reaching 94.9% on GPQA Diamond and large gains on AIME 2025.
-
Overconfidence and Calibration in Medical VQA: Empirical Findings and Hallucination-Aware Mitigation
Empirical study finds overconfidence persists in medical VLMs despite scaling and prompting; post-hoc calibration reduces error while hallucination-aware calibration improves both calibration and AUROC.
-
The Long Delay to Arithmetic Generalization: When Learned Representations Outrun Behavior
Grokking delay on Collatz prediction is a decoder access bottleneck after early encoder structure learning, with numeral base as a strong inductive bias on learnability.
-
Decoding the Critique Mechanism in Large Reasoning Models
By injecting arithmetic mistakes into CoT reasoning, the paper identifies a hidden critique ability in LRMs and extracts a steerable critique vector that enhances self-correction across model scales.
-
TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning
Terminator learns to predict optimal early-exit points in chain-of-thought reasoning by training on the first positions where the model emits its final answer, yielding 14-55% shorter outputs with no accuracy loss.
-
Learning to Stay Safe: Adaptive Regularization Against Safety Degradation during Fine-Tuning
Adaptive regularization guided by training-time safety risk signals from judges or activations prevents safety degradation in fine-tuned language models while preserving utility.
-
Do Reasoning LLMs Refuse What They Infer in Long Contexts?
Long-context LLMs refuse explicit harmful requests but often comply when the same harmful goals must be inferred from distributed fragments in long contexts.
-
FactNet: A Billion-Scale Knowledge Graph for Multilingual Factual Grounding
FactNet is a billion-scale multilingual knowledge graph that links 1.7B Wikidata assertions to 3.01B byte-precise evidence spans from 316 Wikipedia editions, accompanied by a leakage-controlled benchmark suite.
-
Accordion-Thinking: Self-Regulated Step Summaries for Efficient and Readable LLM Reasoning
LLMs learn self-regulated summarization of chain-of-thought steps via RL, allowing compressed Fold inference to reach the same accuracy as exhaustive Unfold mode with far lower token overhead.
-
Can David Beat Goliath? On Multi-Hop Reasoning with Resource-Constrained Agents
David-GRPO improves low-budget RL training for multi-hop QA agents by bootstrapping expert trajectories and converting on-policy partial successes into evidence-coverage signals that increase retrieval depth.
-
"Don't Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo
LLMs exhibit different trade-offs between rule compliance and communicative success across prompting, generation constraints, and representation interventions, but remain substantially weaker than humans at guessing under lexical constraints.
-
Context-Fractured Decomposition Attacks on Tool-Using LLM Agents: Exploiting Artifact Provenance Gaps
Introduces Context-Fractured Decomposition (CFD) attacks exploiting provenance gaps in tool-using LLM agents to raise jailbreak success rates by up to 28.3 percentage points over baselines.
-
DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning
DuMate-DeepResearch introduces a multi-agent deep research system with graph-based planning, recursive execution, and rubric optimization that reports new state-of-the-art scores of 58.03% and 61.95% on two benchmarks.
-
Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment
Cultural zones explain variance in safety ratings beyond demographics across six datasets, with roughly 10% of items identified as culturally sensitive.
-
DynaTree: Dynamic Agentic Retrieval Tree for Time-Sensitive News Retrieval
DynaTree separates offline agentic tree construction from online subtree selection to deliver better recall, ranking, and production survival rates than standard or prior agentic RAG for news retrieval.
-
User-Aware Active Knowledge Acquisition for Emotional Support Dialogue
UKA is a gradient-free active dialogue learning framework using Theory-of-Mind uncertainty estimation to acquire user-aligned conversational knowledge, outperforming baselines in dialogue quality and user alignment across benchmarks.
-
AliMark: Enhancing Robustness of Sentence-Level Watermarking Against Text Paraphrasing
AliMark introduces a two-stage detection strategy with multi-candidate bit sequence alignment to improve robustness of sentence-level text watermarks against paraphrasing attacks.
-
Self-Evolving Spatial Reasoning in Vision Language Models via Geometric Logic Consistency
SAGE adds duality consistency as an auxiliary reward in GRPO training with a dynamic operation pool to improve spatial reasoning robustness and generalization in VLMs.
-
GRC: Unifying Reasoning-Driven Generation, Retrieval and Compression
GRC unifies generation, retrieval, and compression in LLMs via meta latent tokens for single-pass execution with modular flexibility.
-
An Evaluation of Chat Safety Moderations in Roblox
Roblox's automated chat moderation fails to catch numerous unsafe messages involving grooming, sexualization of minors, bullying, violence, self-harm, and sensitive information sharing, with users evading detection through various techniques.
-
Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook
DOVE measures LLM cultural value alignment via a rate-distortion value codebook and unbalanced optimal transport between human and model open-ended text distributions.
-
AI Evaluation Should Require Standardized Item-Level Data Releases
AI benchmark evaluations require standardized item-level data releases as core infrastructure to support validity assessment, demonstrated via the OpenEval archive of 10M responses across 155k items.
-
Calibrate-Then-Act: Cost-Aware Exploration in LLM Agents
Calibrate-Then-Act supplies LLM agents with priors on latent environment states to enable explicit cost-uncertainty reasoning, producing more optimal strategies than standard approaches in retrieval QA and file-reading coding tasks.