MultiHashFormer enables hash-based autoregression in LMs by encoding tokens as multi-hash signatures, outperforming standard Transformers at 100M-3B scales while keeping parameter count constant for multilingual expansion.
Mixed citations
T ruthful QA : Measuring how models mimic human falsehoods
Mixed citation behavior. Most common role is background (50%).
72
Pith papers citing it
460
external citations · Crossref
Background
50% of classified citations
citation-role summary
background 4
dataset 1
method 1
citation-polarity summary
representative citing papers
An empirical study distills a taxonomy of human factual errors from newspaper corrections and shows LLMs achieve only 52% F1 on detection.
AuthorityBench shows citation presence (real or fabricated) increases LLM hallucination rates vs no-citation baseline, strongest for fabricated citations on true claims, with domain variation but negligible venue or author effects.
PhantomBench is a new benchmark of 60K+ non-existent terms showing language models hallucinate at rates up to 86.7 percent even when inputs assume the concepts exist.
Locate-then-edit succeeds at the same early-to-mid MLP locations in masked diffusion models as in autoregressive models, but requires optimization over intermediate partial-mask states to handle multi-token targets.
SuperMemory-VQA provides 4,853 human-verified QA pairs from 52.9 hours of egocentric AI glasses recordings to benchmark AI systems on realistic long-horizon memory tasks including an unanswerable option.
LLMs struggle to associate epistemic markers with stable internal confidence levels across distributions, even under model-centric interpretations, while maintaining somewhat consistent marker rankings.
Commercial AI chatbots reach over 90% multiple-choice accuracy on recent news facts but lose 11-17% in free response and drop to 19-70% on subtle false-premise questions, with retrieval failures causing most errors and clear Anglophone bias.
EvoPref applies NSGA-II evolutionary optimization with archive-based diversity to populations of LoRA adapters, yielding 18% higher preference coverage and 47% lower collapse than gradient descent baselines while matching alignment quality.
LLMs routinely produce unsupported causal stories for personal sensing anomalies, and richer evidence or constrained prompts do not reliably eliminate this epistemic overreach.
A new paired-prompt protocol reveals alignment-pipeline-specific heterogeneity in how open-weight LLMs respond to evaluation versus deployment framings.
SemGrad measures LLM uncertainty via gradients in semantic space using a Semantic Preservation Score to select embeddings, with HybridGrad combining it with parameter gradients to outperform sampling-based baselines especially when multiple responses are valid.
SENECA uses a novel self-consistent missing mass calculation to improve discrete entropy estimates in small-sample regimes and outperforms alternatives in numerical tests.
VLMs as judges exhibit informativeness bias by favoring detailed but image-inconsistent answers; BIRCH mitigates it by first correcting answers against the image, reducing bias up to 17% and improving performance up to 9.8%.
Alignment of vision-language models with human V1-V3 early visual cortex negatively predicts resistance to sycophantic gaslighting attacks.
Agreeableness in AI personas reliably predicts sycophantic behavior in 9 of 13 tested language models.
CapTrack shows post-training causes drift beyond facts, with instruction fine-tuning producing stronger behavioral changes than preference optimization across model families.
Norm-Anchor Scaling breaks the norm-feedback loop in sequential LLM editing by anchoring value vectors to original norms, improving long-run performance by 72.2% and extending the editing horizon over 4x.
LLM unlearning is reframed as inadvertently installing backdoor triggers on forget-tokens; Random Noise Augmentation is introduced as a defense that improves robustness with theoretical guarantees.
GAIA benchmark shows humans at 92% accuracy on simple real-world questions far outperform current AI systems at 15%, proposing this gap as a key milestone for general AI.
Chain-of-Thought reasoning in LLMs is often unfaithful, with models relying on it variably by task and less so as models scale larger.
Pruning MoE LLMs according to the stationary distribution of a Markov chain over (layer, expert) routing transitions retains more task performance than local importance heuristics, with up to ~3.5% relative gains over the strongest baseline at 50% compression.
Composable LoRA adapters can amplify or suppress OCEAN traits in LLMs, combine approximately additively, preserve moderate-scale capability, and move safety-relevant behaviours.
Hallucinations arise from biased latent inference paths rather than missing knowledge, demonstrated via a new diagnostic testbed TrapQA that isolates task-retrieval and key-selection biases.
citing papers explorer
-
MultiHashFormer: Hash-based Generative Language Models
MultiHashFormer enables hash-based autoregression in LMs by encoding tokens as multi-hash signatures, outperforming standard Transformers at 100M-3B scales while keeping parameter count constant for multilingual expansion.
-
An Empirical Analysis of Factual Errors in Human-Written Text and its Application
An empirical study distills a taxonomy of human factual errors from newspaper corrections and shows LLMs achieve only 52% F1 on detection.
-
Authority, Truth, and Citation Bias: A Large-Scale Multi-Domain Benchmark for Studying Epistemic Susceptibility in Large Language Models
AuthorityBench shows citation presence (real or fabricated) increases LLM hallucination rates vs no-citation baseline, strongest for fabricated citations on true claims, with domain variation but negligible venue or author effects.
-
PhantomBench: Benchmarking the Non-existential Threat of Language Models
PhantomBench is a new benchmark of 60K+ non-existent terms showing language models hallucinate at rates up to 86.7 percent even when inputs assume the concepts exist.
-
Knowledge Editing in Masked Diffusion Language Models
Locate-then-edit succeeds at the same early-to-mid MLP locations in masked diffusion models as in autoregressive models, but requires optimization over intermediate partial-mask states to handle multi-token targets.
-
SuperMemory-VQA: An Egocentric Visual Question-Answering Benchmark for Long-Horizon Memory
SuperMemory-VQA provides 4,853 human-verified QA pairs from 52.9 hours of egocentric AI glasses recordings to benchmark AI systems on realistic long-horizon memory tasks including an unanswerable option.
-
Can LLMs Use Linguistic Uncertainty Markers to Reliably Reflect Intrinsic Confidence?
LLMs struggle to associate epistemic markers with stable internal confidence levels across distributions, even under model-centric interpretations, while maintaining somewhat consistent marker rankings.
-
Evaluating Commercial AI Chatbots as News Intermediaries
Commercial AI chatbots reach over 90% multiple-choice accuracy on recent news facts but lose 11-17% in free response and drop to 19-70% on subtle false-premise questions, with retrieval failures causing most errors and clear Anglophone bias.
-
EvoPref: Multi-Objective Evolutionary Optimization Discovers Diverse LLM Alignments Beyond Gradient Descent
EvoPref applies NSGA-II evolutionary optimization with archive-based diversity to populations of LoRA adapters, yielding 18% higher preference coverage and 47% lower collapse than gradient descent baselines while matching alignment quality.
-
Causal Stories from Sensor Traces: Auditing Epistemic Overreach in LLM-Generated Personal Sensing Explanations
LLMs routinely produce unsupported causal stories for personal sensing anomalies, and richer evidence or constrained prompts do not reliably eliminate this epistemic overreach.
-
Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity
A new paired-prompt protocol reveals alignment-pipeline-specific heterogeneity in how open-weight LLMs respond to evaluation versus deployment framings.
-
Gradients with Respect to Semantics Preserving Embeddings Tell the Uncertainty of Large Language Models
SemGrad measures LLM uncertainty via gradients in semantic space using a Semantic Preservation Score to select embeddings, with HybridGrad combining it with parameter gradients to outperform sampling-based baselines especially when multiple responses are valid.
-
SENECA: Small-Sample Discrete Entropy Estimation via Self-Consistent Missing Mass
SENECA uses a novel self-consistent missing mass calculation to improve discrete entropy estimates in small-sample regimes and outperforms alternatives in numerical tests.
-
When Vision-Language Models Judge Without Seeing: Exposing Informativeness Bias
VLMs as judges exhibit informativeness bias by favoring detailed but image-inconsistent answers; BIRCH mitigates it by first correcting answers against the image, reducing bias up to 17% and improving performance up to 9.8%.
-
Gaslight, Gatekeep, V1-V3: Early Visual Cortex Alignment Shields Vision-Language Models from Sycophantic Manipulation
Alignment of vision-language models with human V1-V3 early visual cortex negatively predicts resistance to sycophantic gaslighting attacks.
-
Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models
Agreeableness in AI personas reliably predicts sycophantic behavior in 9 of 13 tested language models.
-
CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training
CapTrack shows post-training causes drift beyond facts, with instruction fine-tuning producing stronger behavioral changes than preference optimization across model families.
-
Norm Anchors Make Model Edits Last
Norm-Anchor Scaling breaks the norm-feedback loop in sequential LLM editing by anchoring value vectors to original norms, improving long-run performance by 72.2% and extending the editing horizon over 4x.
-
Improving LLM Unlearning Robustness via Random Perturbations
LLM unlearning is reframed as inadvertently installing backdoor triggers on forget-tokens; Random Noise Augmentation is introduced as a defense that improves robustness with theoretical guarantees.
-
GAIA: a benchmark for General AI Assistants
GAIA benchmark shows humans at 92% accuracy on simple real-world questions far outperform current AI systems at 15%, proposing this gap as a key milestone for general AI.
-
Measuring Faithfulness in Chain-of-Thought Reasoning
Chain-of-Thought reasoning in LLMs is often unfaithful, with models relying on it variably by task and less so as models scale larger.
-
It Takes a MAESTRO To Prune Bad Experts
Pruning MoE LLMs according to the stationary distribution of a Markov chain over (layer, expert) routing transitions retains more task performance than local importance heuristics, with up to ~3.5% relative gains over the strongest baseline at 50% compression.
-
Persona Cartography: Charting Language Model Personality Traits in Weight Space
Composable LoRA adapters can amplify or suppress OCEAN traits in LLMs, combine approximately additively, preserve moderate-scale capability, and move safety-relevant behaviours.
-
Understanding Why Language Models Hallucinate: Testing Reasoning Against Priors
Hallucinations arise from biased latent inference paths rather than missing knowledge, demonstrated via a new diagnostic testbed TrapQA that isolates task-retrieval and key-selection biases.
-
Truth or Sophistry? LoFa: A Benchmark for LLM Robustness Against Logical Fallacies
LoFa is a new benchmark and LFR@k metric for measuring LLM resistance to sustained logical fallacy attacks via generated question-argument pairs and debate simulations.
-
What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs
Proposes SCSuff metric for evaluating LLM explanation sufficiency via model-generated alternative inputs, showing explanations are typically insufficient and predictable from hidden states.
-
A Risk Decomposition Framework for Pre-Hoc Fine-Tuning Prediction
Formulates pre-hoc fine-tuning prediction as stochastic estimation, proves lower bound on optimization variance decay rate, and introduces a three-regime predictability phase diagram.
-
Sch\"utzen: Evaluating LLM Safety in Bulgarian and German Contexts
Schützen is a German-Bulgarian LLM safety dataset showing pronounced cross-language differences in model safety behavior.
-
Quality Is Not a Safety Proxy Under Quantization
Across 51 quantized checkpoints, quality metrics fail to predict safety drops in 36 pairings and 10 hidden-danger cases, while a new RTSI screen routes all 10 dangerous rows to testing at matched bucket size.
-
GIScholarBench: Benchmarking LLM Overconfidence in GIS Research
GIScholarBench shows LLMs exhibit consistent overconfidence across three scholarly tasks in GIS, with different manifestations in factual retrieval, citation expansion, and idea generation.
-
Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation
Post-hoc model-based compression of reasoning traces cuts training tokens to 12-30% and speeds training 2-7.6x while retaining up to 96% of raw-trace accuracy, though raw traces remain superior at every scale.
-
TENP: Trapezoidal Expert Neuron Pruning For Mixture-of-Experts
TENP applies trapezoidal expert-neuron pruning to MoE models, retaining key experts while pruning others via projected neuron contribution, yielding only 1-point accuracy drop at 40% sparsity on DeepSeek with 10% code-generation gain.
-
Large Language Models Are Overconfident in Their Own Responses
Instruction-tuned LLMs exhibit an ownership bias, assigning up to 26% higher confidence to their own responses than identical user-provided answers; reframing the answer as user input during elicitation reduces overconfidence by up to 26%.
-
Investigating and Alleviating Harm Amplification in LLM Interactions
Presents HarmAmp benchmark for multi-turn harm amplification in LLMs and TrajSafe proactive monitor that reduces harm while keeping low over-refusal and preserving capabilities.
-
Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time
RCA is a training-free module that boosts input context signal strength in the residual stream of LLMs by orthogonal decoupling of attention routing from value magnitude.
-
Not How Many, But Which: Parameter Placement in Low-Rank Adaptation
Gradient-informed placement of LoRA parameters recovers full performance under GRPO while random placement does not, due to differences in gradient rank and stability across training regimes.
-
A CAP-like Trilemma for Large Language Models: Correctness, Non-bias, and Utility under Semantic Underdetermination
Under semantic underdetermination, LLMs cannot always guarantee strong correctness, strict non-bias, and high utility at once.
-
PRISM: A Geometric Risk Bound that Decomposes Drift into Scale, Shape, and Head
PRISM supplies a geometric upper bound on LLM variant risk that splits drift into scale, shape, and head axes and doubles as a differentiable regularizer against forgetting.
-
Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models
Dynamic Boundary Evaluation locates each LLM's performance boundary at ~50% pass probability via a calibrated item bank and Skill-Guided Boundary Search algorithm to enable unified, adaptive evaluations across safety, capability, and truthfulness.
-
Debiasing Reward Models via Causally Motivated Inference-Time Intervention
Neuron-level inference-time intervention reduces multiple biases in reward models, enabling 2B and 7B models to match 70B performance on LLM alignment benchmarks without trade-offs.
-
Layerwise Convergence Fingerprints for Runtime Misbehavior Detection in Large Language Models
LCF detects multiple LLM runtime threats by computing aggregated diagonal Mahalanobis distances on layer-wise hidden-state differences, calibrated on clean examples, achieving high detection rates with low overhead across several model architectures.
-
Architecture Determines Observability of Transformers
Architecture and training determine whether transformers retain a readable internal signal that lets activation monitors catch errors missed by output confidence.
-
ReFACT: A Benchmark for Scientific Confabulation Detection with Positional Error Annotations
ReFACT benchmark reveals LLMs show a persistent salient distractor failure mode where 61% of incorrect error span predictions are semantically unrelated to actual errors, persisting across model sizes, and comparative judgment yields lower F1 than independent detection.
-
Token-Level Density-Based Uncertainty Quantification Methods for Eliciting Truthfulness of Large Language Models
Adapts multi-layer token-level Mahalanobis distance with supervised linear regression to yield improved uncertainty scores for LLM truthfulness tasks.
-
Unconditional Truthfulness: Learning Unconditional Uncertainty of Large Language Models
A regression model using attention features and recurrent uncertainty scores improves selective generation in LLMs over unsupervised and supervised baselines on ten datasets and three models.
-
Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models
Empirical analysis shows scaling inference compute via strategies like tree search can be more efficient than scaling model parameters, with 7B models plus novel search outperforming 34B models.
-
Towards Understanding Sycophancy in Language Models
Sycophancy is prevalent in state-of-the-art AI assistants and is likely driven in part by human preferences that favor agreement over truthfulness.
-
The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
At sufficient scale, LLMs linearly represent the truth value of factual statements, as shown by visualizations, cross-dataset generalization, and causal interventions that flip truth judgments.
-
Measuring Progress on Scalable Oversight for Large Language Models
Humans chatting with an unreliable LLM assistant outperform both the model alone and unaided humans on MMLU and time-limited QuALITY tasks.
-
Large Language Models Are Human-Level Prompt Engineers
APE generates instruction candidates via LLM and selects the best by zero-shot performance of a second LLM, matching or beating human prompts on 19 of 24 NLP tasks.