Papers for

automated content moderation teams

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Combining human and AI feedback to better rank language models

Prediction-Powered LLM Ranking with Adaptively Collected Human Feedback

Abstract: Ranking large language models (LLMs) through human evaluations is essential for downstream decision-making. Because human feedback is expensive, evaluation platforms often collect it adaptively, prioritizing informative model comparisons using previously observed feedback. To further improve precision with limited human feedback, evaluators can draw on abundant unrated responses and relatively inexpensive LLM judges; however, judge bias and adaptive dependence challenge standard i.i.d. inference frameworks. We propose Prediction-Powered Ranking (PP-Rank), a framework that combines LLM-judge predictions with human feedback to improve precision while providing valid inference under adaptive sampling. Data-driven prediction weights minimize the asymptotic covariance trace within the proposed class. In theory, we establish asymptotic normality for general evaluation targets and simultaneous ranking inference. A joint martingale analysis handles dependence between adaptive human feedback and auxiliary data. We further derive a rule for reducing exploration that prioritizes lower-cost comparisons while matching a conservative variance bound, and characterize allocation limits for regular $\sqrt{T}$-rate estimation of pairwise means. Semi-synthetic experiments on UltraFeedback-Extended show improvements over human-only baselines in score estimation, confidence interval width, and ranking certification. Synthetic experiments demonstrate robustness to judge quality and adaptive allocation, while showing that PP-Rank can substantially reduce evaluation costs without sacrificing precision.

Mon 9 NovMachine Learning
The gist
Ranking language models by using human feedback helps decide which model works best, but getting lots of human evaluations is expensive. The authors propose a method called PP-Rank that mixes human feedback with predictions from AI judges to improve ranking accuracy without needing many human reviews. Their approach handles tricky dependencies from collecting data step-by-step and adjusts how to collect feedback more efficiently. Experiments show this method can save evaluation costs while keeping results trustworthy.
Open → 2611.13186v1

Agora method improves aggregating opinions of large language model judges

AGORA: Reliability-Aware Aggregation of LLM Judges via Item-Response Theory

Abstract: LLM-as-a-judge is a common substitute for costly human evaluation, and aggregating several judges is the standard remedy for any single judge's bias. Judge pools mix judges of very different reliability, while the human-labeled sets that validate them hold a few hundred items or fewer. In this regime the per-item mean weights weak judges as heavily as strong ones, and confusion-table models such as Dawid-Skene must estimate a per-judge error table from too few items. We propose AGORA (Aggregation Of Reliability-weighted Assessments), an item-response-theory aggregator that places every item on one latent quality scale and describes each judge by a sensitivity parameter and ordered rating thresholds, so judge reliability is learned from all items jointly and without human labels. Across three benchmarks, AGORA matches or exceeds Dawid-Skene on large evaluation sets, and its advantage widens as the set shrinks toward a few hundred items. It remains accurate as pools grow more uneven, a label-free dispersion statistic indicates in advance when trust weighting helps, and its learned trust weights track per-judge human validity.

Tue 3 NovArtificial IntelligenceComputation and Language
The gist
When many AI models called large language models (LLMs) act as judges for tasks usually done by humans, their opinions can vary in accuracy. The authors propose a way to weight each AI judge’s opinion based on reliability, learned without needing examples judged by humans. Their method, AGORA, puts every task on a quality scale and figures out how sensitive and trustworthy each judge is by looking at all tasks together. Tests show AGORA works better especially when the number of tasks is small or judges differ a lot in skill.
Open → 2611.05556v1

Vision language model ranking improves by detecting cue reliance shifts

Shift-Reliance Alignment: Ranking Vision-Language Models by the Cues They Trust

Abstract: Vision-language models (VLMs) are usually chosen from leaderboards, which give one ranking for every application. On questions about objects that a scene suggests but does not contain, however, the model with the lowest error on the POPE hallucination benchmark is not the one with the lowest error on such questions. For yes/no questions about object presence, we show that the best model for a target domain can be predicted without running any candidate on that domain, given 650 to 1,200 labeled target items that calibrate a detector. We measure each model's reliance on a cue, a feature such as scene context or text in the image, by editing 100 labeled source images so that the cue suggests the wrong answer, and comparing with a matched control edit. Reliance predicts which VLMs are misled by physical sticker attacks (Spearman .93 on nine models not used to design the score) and which hallucinate on unedited COCO photographs, beyond what their tendency to answer yes explains. Shift-Reliance Alignment (SRA) scores a model by its source error on the labeled source questions plus its reliance multiplied by the shift. The shift is the proportion of target questions on which the cue misleads, and a detector estimates it on the unlabeled target images. SRA selects the model with the lowest score. In pre-registered tests on new images where most questions mislead, it lowered the error of the selected model from 30.4% under selection by source error to 18.8% on Objects365 and OpenImages but, against its prediction, not on COCO. On CueDial, a benchmark we release that varies how often a cue misleads, selection by model confidence falls more than 40 percentage points behind the best model when overlaid text denies a visible object, and SRA stays within 3 points.

Thu 15 OctComputer Vision and Pattern RecognitionMachine Learning
The gist
Vision-language models are often ranked by their overall accuracy, but this doesn’t always predict how well they handle tricky questions about objects that might be suggested but aren’t actually present. The authors found a way to predict which model will perform best in a new setting by measuring how much each model relies on certain clues, like background context or text in images. They created a method called Shift-Reliance Alignment that combines the model’s original error with how much misleading cues change its answers, helping select a better model without testing all options extensively. Their approach better reduced errors on some image sets but not all, showing it can help choose models that avoid certain mistakes.
Open → 2610.22610v1

Dualview improves multipoint text scoring with fewer parameters

DUALVIEW: Leveraging Broad and Output-Specific Evidence for LLM-based Multi-Output Text Regression

Abstract: While work on LLM-based regression has largely focused on scalar prediction, multi-output text regression remains less explored. We present DualView, a parameter-efficient approach to leverage LLMs for multi-output text regression. \method{} first extracts broad and output-specific representations from a frozen LLM, then uses a dual-branch architecture to produce the full score vector. Across five heterogeneous regression tasks and eight LLMs (up to 122B), DualView outperforms seven supervised regression baselines while requiring 82--1842 times fewer fitted parameters than the three strongest baselines. These improvements extend across individual output dimensions and persist even with limited training data. Notably, DualView using a smaller open-weight LLM can outperform unsupervised prediction with Gemini 3.1 Pro. Controlled ablations show that broad and output-specific evidence provide complementary predictive benefits, while lightweight shared predictors achieve strong performance with fewer trainable parameters.

Tue 13 OctComputation and Language
The gist
Assigning multiple scores to text using large language models is hard and usually needs lots of training. The authors created DualView, a method that uses both general and specific information from a large language model without changing its core. DualView produces multiple related outputs efficiently and performs better than other training methods, even when data is limited. This approach works well with different models and tasks.
Open → 2610.19206v1

Label collapse undermines reliability of diffusion language model predictions

Label Collapse in Diffusion Language Model In-Context Learning

Abstract: In-context learning (ICL) applies large language models to new tasks by prepending a few labeled demonstrations to a query. Prior work has studied how demonstration ordering affects ICL, but the interaction between label arrangement patterns and model biases has not been systematically examined. We define label collapse as the failure mode in which the model generates query-independent labels, and show that it is architecturally universal at the tested 8-9B scale: five models spanning autoregressive (Llama-3, Qwen3.5) and discrete diffusion (LLaDA, iLLaDA, LLaDA-Instruct) families all exhibit label collapse even under the class-balanced alternating arrangement, with some label word pairs collapsing in virtually every tested condition and others flipping their collapse direction with arrangement. Critically, accuracy alone cannot detect this collapse: models can achieve near-chance accuracy while generating query-independent labels, a score indistinguishable from balanced performance, so the failure can persist undetected in deployment. We measure separation, the log-odds of the two label words at the query position (the quantity that label-bias calibration operates on), as a complementary diagnostic that exposes collapse hidden beneath surface accuracy, together with a measurement protocol that guards label extraction against tokenizer and formatting artifacts. For autoregressive models, collapse persists where the majority-label-bias account does not apply, and arrangement patterns can reverse the collapse direction; for diffusion models, we present a systematic evaluation of ICL failure modes, complementing contemporaneous studies of dLLM context comprehension and demonstration-order sensitivity, and identify open questions about how iterative denoising interacts with label biases.

Mon 12 OctComputation and Language
The gist
In-context learning lets large language models solve new tasks by showing examples before a question. The authors find a problem called label collapse where models ignore the question and repeat the same label no matter what. This happens across different types of models and setups, and normal accuracy scores don't reveal this issue. They suggest a new way to detect label collapse by measuring how the model differentiates between label options, helping catch this hidden failure.
Open → 2610.16839v1

VisExMEM verifies image captions with claim-level visual evidence

VisExMEM: Visual Evidence-Based Explainable Multimodal Evaluation Metric

Abstract: Current image-text evaluation often relies on a global score, which gives limited insight into which semantic details are supported by the image. Explainable metrics provide rationales, but the visual evidence supporting those rationales can still be difficult to verify. We introduce VisExMEM, a reference-free evaluator that decomposes captions into atomic claims, localizes visual evidence, verifies each claim, and averages claim-level support into a global score. Explanations use the same evidence and decisions. We train Qwen3.5-based 4B and 9B models on 100K samples with multi-task SFT, counterfactual refinement, and evidence-aware GDPO. Across eight benchmarks, VisExMEM-9B achieves the highest descriptive composite among open-source evaluators in our comparison, exceeding VQAScore by 11.7 points while remaining competitive with GPT-5.6 Sol. We also introduce AGITA-Bench, containing 600 human-annotated image-text pairs and 8.7K atomic claims for global and claim-level evaluation. For untuned Qwen3.5 and GPT-5.6 Sol controls, decomposed prompting raises the same composite by 3.6-5.1 points relative to holistic prompting. Human evaluation shows that VisExMEM explanations outperform prior explainable metrics and approach GPT-5.6 Sol.

Mon 12 OctComputer Vision and Pattern RecognitionArtificial IntelligenceComputation and Language
The gist
Evaluating how well captions describe images is usually done with one overall score, which doesn't show which parts of the caption are correct. The authors created VisExMEM, a tool that breaks down captions into small claims, finds the matching visual evidence in the image, and checks if each claim is supported. This approach gives a more detailed and explainable evaluation. Their models outperform existing open-source evaluators and provide explanations that people find clearer.
Open → 2610.16441v1

Stateless service gives repeatable text changes preserving original meaning

perturber: A Stateless Service for Semantics-Preserving Text Perturbations

Abstract: Large language model (LLM) evaluations on clean benchmarks overstate model robustness: real inputs arrive perturbed, and these meaning-preserving variants routinely move model answers. Studying this systematically requires applying perturbations one at a time, reproducibly, and identically wherever they run, yet existing tools couple them to a benchmarking stack rather than packaging them as a reusable transform layer. We present perturber, a stateless service and Python library that applies exactly one semantics-preserving perturbation per call, drawn from a first-party catalog of 68 perturbations across 11 families. It turns on a single primitive, with composition left to the caller chaining calls, so the service and the library produce identical output by construction. We evaluate the tool itself, its correctness, reproducibility, and cost, rather than any model's robustness, measuring meaning-preservation per perturbation rather than assuming it. The demonstration adds a live playground, an introspectable catalog for benchmark selection, and a one-command reproducible evaluation.

Thu 8 OctComputation and Language
The gist
Evaluations of language AI often miss how small changes in sentences can confuse these systems even when the meaning stays the same. The authors created perturber, a tool and service that makes one small, meaning-preserving change to a piece of text each time it runs, ensuring consistent results no matter where or when it's used. This helps others test how robust language models really are by applying known, measured text changes one at a time. They also provide a live demo and library to use these changes easily for evaluation.
Open → 2610.12667v1

Controlled fusion of audio video and text improves decision making with limited data

Controlled Acquisition and Abstention in Three-Channel Score Conflicts

Abstract: When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

Wed 7 OctMachine LearningMultimedia
The gist
When audio, video, and text information don’t agree, just measuring accuracy doesn’t tell us when to look for more information or decide to not answer. The authors created a setup where a system sees two scores, can pay to get the third, or choose to abstain from deciding. They found that a smart rule using thresholds picks when to get more info or abstain better than always answering or always asking for everything, depending on the situation. This helps show that choosing what data to use before asking for more is important and depends on how much data you want to spend and what’s already seen.
Open → 2610.10808v1

Vision language models reveal new clues to spot adversarial images

Detecting Adversarial Images through Response Profiles of Vision-Language Models

Abstract: Adversarial perturbations can alter the predictions of frozen vision-language models (VLMs) while leaving their confidence and image--text similarity patterns seemingly plausible. We investigate whether we can identify adversarial inputs based on the broader way an image interacts with a collection of general semantic prompts. Our detector summarizes these responses using category-level statistics, relationships among prompts, deviations from clean reference distributions, and stability under weak image transformations, producing a compact response profile that is classified by a lightweight model while the VLM remains fixed. We evaluate the approach on multiple public image datasets, several CLIP-style visual backbones, and a range of gradient-based, optimization-based, automated, and spatial attacks. The detector achieves strong discrimination in attack-specific settings and retains substantial performance when evaluated on attacks not seen during training. Under a controlled detector-specific protocol, the response-profile representation outperforms the evaluated embedding-geometry baselines. Additional analyses show that the feature groups provide complementary information and that the method remains effective under variations in the prompt configuration. We also examine inference cost and performance against detector-aware adaptive attacks. Overall, the results indicate that response patterns across semantic prompts provide a useful complementary signal for adversarial image detection in frozen VLMs.

Wed 7 OctComputer Vision and Pattern Recognition
The gist
Adversarial images are specially altered pictures designed to trick AI models into making wrong predictions without looking suspicious. The authors found that by looking at how a frozen vision-language model responds to many different text prompts, they can detect these tricky images better. Their method creates a summary of these responses and uses a simple classifier to decide if an image is adversarial. This technique works well on various attacks and even when facing attacks it hasn't seen before.
Open → 2610.10436v1

Llm judges adapt to human preferences despite position bias

DIAL: Position-Debiased LLM Judges with Adaptive Human Preference Calibration

Abstract: Large language models (LLMs) as a judge enable scalable evaluation, but their judgments can be sensitive to response order and, even after removing such position effects, can still diverge systematically from human preferences.We introduce DIAL, a unified framework that combines abundant LLM comparisons with limited human comparisons to separate judge-specific position effects, learn shared structure in position-debiased LLM preferences, and adaptively calibrate that structure toward the human preference target. Theoretically, we study three aspects of DIAL: (i) identification of latent LLM preferences, position effects, and human calibration; (ii) adaptive estimation that balances LLM anchoring against limited human evidence; and (iii) fixed-weight uncertainty quantification for the calibrated human preference. Empirically, we evaluate position debiasing and human alignment separately in controlled simulations and on three human-preference benchmarks, showing that DIAL remains robust to unbalanced response order, achieves strong human-aligned rankings with limited labels, and adapts toward human evidence when LLM information is imperfect. Our real-data study collects over 410K judgments from 21 LLM judges in both display orders, providing a resource for future studies of LLM-judge bias, heterogeneity, and human alignment.

Fri 25 SeptArtificial Intelligence
The gist
Large language models can judge other AI responses but their opinions can be influenced by the order in which results are shown and differ from human choices. The authors introduce a method called DIAL that uses AI comparisons along with a small amount of human feedback to remove these biases and adjust the AI judgments to better match human preferences. They test this approach with many AI judges and human-labeled benchmarks and find DIAL creates rankings closer to what people want, even when the AI information isn’t perfect.
Open → 2609.31215v1

Visual question answering improves safety using evidence based agent collaboration

A$^2$Safe: Counterfactual Evidence-Aligned Adaptive Agent Collaboration for Safe and Effective Visual Question Answering

Abstract: Visual Question Answering (VQA) with Multimodal Large Language Models (MLLMs) requires not only producing safe and effective responses, but also grounding safety decisions in the multimodal evidence that determines risk. Recent safety-alignment methods improve refusal behavior and contextual risk awareness, yet correct safety outcomes may still rely on superficial textual or visual correlations, particularly when risk emerges from interactions between individually benign image and question content. To address this issue, we propose A$^2$Safe, a counterfactual evidence-aligned adaptive agent collaboration framework for safe and effective VQA. A$^2$Safe organizes localized visual observations, textual intent, and cross-modal risk relations through a Grounded Safety Evidence Board, making the basis of safety decisions explicit. Counterfactual safety evidence alignment enforces invariance to safety-irrelevant changes while requiring appropriate safety-state and response-mode transitions when risk-critical evidence is minimally altered. The resulting evidence state further supports adaptive collaboration, enabling direct answering when grounded evidence is sufficient and invoking policy critique and response revision when evidence is risky, uncertain, or conflicting. Under complementary safety-critical and general VQA protocols, A$^2$Safe achieves a 95.72 SIUO safety score, reduces the benign refusal rate on MOSSBench to 14.67%, and maintains an average general VQA score of 78.34 with 27.8% token overhead. These results support counterfactual evidence-aligned adaptive collaboration for safe and effective multimodal question answering.

Mon 21 SeptComputer Vision and Pattern Recognition
The gist
Visual Question Answering (VQA) models that understand images and text need to give safe and accurate answers, especially when the question and image together might create risk. The authors propose a new approach called A²Safe that carefully tracks and aligns the evidence in both the image and question to make clear safety decisions. This method adapts how it responds, either directly answering or questioning the evidence when uncertain or risky. Their approach improves safety metrics and reduces unnecessary refusals while keeping overall answer quality high.
Open → 2609.24098v1

Flow matching stability improved by budgeting per-step variance in image generation

$λ$-Controlled GRPO: Turning Flow-Matching Ratio Instability into a Budgeted Resource

Abstract: Reinforcement learning is increasingly used to align image generators with reward signals, and Flow-GRPO recently extended this paradigm to flow-matching models by treating the denoising sampler as a stochastic policy that can be optimized from reward feedback. Training in this setting is unstable in a way specific to multi-step denoising: the policy update changes systematically across denoising steps, with importance ratios drifting below one, becoming increasingly dispersed, clipping at different rates, and leaving fewer usable samples late in training. Prior work treats these effects as separate failure modes and addresses each with a hand-tuned stabilizer. We show instead that they arise from a single per-step quantity, which we call path variance. This quantity is determined exactly by the sampler's Gaussian transition kernel and can be estimated cheaply during training. This reframes instability as a resource that can be measured and budgeted rather than a collection of symptoms to repair. Our method, $λ$-Controlled GRPO, calibrates importance-ratio behavior from this predicted law rather than from noisy empirical statistics, and allocates gradient effort across denoising steps according to their predicted cost. The two scales governing the update are fixed by standard policy choices rather than introduced as free tuning parameters. On a text-to-image model under two reward settings, rendering difficult target text scored by optical character recognition and matching human preferences scored by a preference model, $λ$-Controlled GRPO improves both text accuracy and preference reward over the strongest empirical stabilizer. It also keeps late-step path variance within its intended budget, precisely where the baseline systematically overshoots. The result is a Flow-GRPO update calibrated by its own transition law rather than stabilized after instability appears.

Fri 18 SeptMachine Learning
The gist
Training AI models that create images based on rewards can become unstable during the step-by-step process of image generation. The authors identify that this instability comes from a single measurable factor called path variance, linked to the way the model samples data. Instead of trying to fix many symptoms separately, they control this path variance using a new method that predicts and budgets it throughout training. This approach helps the model better understand tricky text instructions and align with human preferences more reliably without relying on guesswork.
Open → 2609.22041v1