Daily Papers

The pass-all-test binary reward has become a standard recipe for training software engineering agents, but its sparsity eliminates difference of partially correct patches. The test pass-rate is a natural alternative. However, we find that it is not a reliable indicator of patch quality. A functionality being tested can be covered by either one challenging test or one hundred trivial tests, where the majority might be already satisfied pass-to-pass tests. We introduce CodeLift, a reinforcement learning method that dynamically calibrates pass-rate rewards, turning the test suite into an adaptive curriculum. At each iteration, CodeLift groups tests that kill the same patches and patches that pass the same tests, then gives more weight to test groups that fewer patch groups pass. This reduces the impact of majority simple tests and focuses more on what is still challenging. Within each rollout, CodeLift applies the same scoring rule to repository snapshots, assigning credit to useful actions even when later ones introduce regressions. Experiments on SWE benchmarks show that CodeLift consistently improves pass@1/2/4/8 over binary and pass-rate baselines. Analyses show that CodeLift provides dense, stable learning signals and unlocks the inherent curriculum in the test suite.

A critical challenge in recommender systems is to establish reliable relationships between offline and online metrics that predict real-world performance. Motivated by recent advances in Pareto front approximation, we introduce a pragmatic strategy for identifying offline metrics that align with online impact. A key advantage of this approach is its ability to simultaneously serve multiple test groups, each with distinct offline performance metrics, in an online experiment controlled by a single model. The method is model-agnostic for systems with a neural network backbone, enabling broad applicability across architectures and domains. We validate the strategy through a large-scale online experiment in the field of session-based recommender systems on the OTTO e-commerce platform. The online experiment identifies significant alignments between offline metrics and real-word click-through rate, post-click conversion rate and units sold. Our strategy provides industry practitioners with a valuable tool for understanding offline-to-online metric relationships and making informed, data-driven decisions.

Streaming spoken agents may produce the correct final action after acting too early. Final-turn scores do not reveal whether each observed speech prefix supports an exposed action. We introduce the Partial Speech Action Contract for Turn Taking in Speech Language Models (PACT-SLM), a controlled test that assigns the first valid action time and evaluates both action identity and timing. In the primary test, 80 paired contrast groups from four held-out semantic families yield 1,600 prefix predictions across clean and 15 dB noise renderings. Using source-utterance semantic targets rather than counterbalanced branch codes, WavLM Base Plus reaches 26.03% pooled post-onset semantic-label accuracy (95% group-bootstrap interval [22.14%, 29.68%]), exposes an action on 18.99% of pre-onset prefixes, and predicts 5.94% of complete trajectories exactly. Its pooled label score is at the 96th percentile of 100 within-prefix label permutations, below the 97.5th-percentile reference (26.73%). It exceeds matched text, scalar-acoustic, and shuffled-representation probes in pooled post-onset label accuracy. Elapsed time is more onset-exact (36.25% versus 23.13%) but less accurate about action identity (9.92% versus 26.03%). These results motivate separate measurement of action identity and onset timing in partial-speech evaluations.

Bias in Large Language Models (LLMs) significantly undermines their reliability and fairness. We focus on a common form of bias: when two reference concepts in the model's concept space, such as sentiment polarities (e.g., "positive" and "negative"), are asymmetrically correlated with a third, target concept, such as a reviewing aspect, the model exhibits unintended bias. For instance, the understanding of "food" should not skew toward any particular sentiment. Existing bias evaluation methods assess behavioral differences of LLMs by constructing labeled data for different social groups and measuring model responses across them, a process that requires substantial human effort and captures only a limited set of social concepts. To overcome these limitations, we propose BiasLens, a test-set-free bias analysis framework based on the structure of the model's vector space. BiasLens combines Concept Activation Vectors (CAVs) with Sparse Autoencoders (SAEs) to extract interpretable concept representations, and quantifies bias by measuring the variation in representational similarity between the target concept and each of the reference concepts. Even without labeled data, BiasLens shows strong agreement with traditional bias evaluation metrics (Spearman correlation r > 0.85). Moreover, BiasLens reveals forms of bias that are difficult to detect using existing methods. For example, in simulated clinical scenarios, a patient's insurance status can cause the LLM to produce biased diagnostic assessments. Overall, BiasLens offers a scalable, interpretable, and efficient paradigm for bias discovery, paving the way for improving fairness and transparency in LLMs.

Machine learning methods strive to acquire a robust model during the training process that can effectively generalize to test samples, even in the presence of distribution shifts. However, these methods often suffer from performance degradation due to unknown test distributions. Test-time adaptation (TTA), an emerging paradigm, has the potential to adapt a pre-trained model to unlabeled data during testing, before making predictions. Recent progress in this paradigm has highlighted the significant benefits of using unlabeled data to train self-adapted models prior to inference. In this survey, we categorize TTA into several distinct groups based on the form of test data, namely, test-time domain adaptation, test-time batch adaptation, and online test-time adaptation. For each category, we provide a comprehensive taxonomy of advanced algorithms and discuss various learning scenarios. Furthermore, we analyze relevant applications of TTA and discuss open challenges and promising areas for future research. For a comprehensive list of TTA methods, kindly refer to https://github.com/tim-learn/awesome-test-time-adaptation.

Algorithmic decision-making has become deeply ingrained in many domains, yet biases in machine learning models can still produce discriminatory outcomes, often harming unprivileged groups. Achieving fair classification is inherently challenging, requiring a careful balance between predictive performance and ethical considerations. We present FairTTTS, a novel post-processing bias mitigation method inspired by the Tree Test Time Simulation (TTTS) method. Originally developed to enhance accuracy and robustness against adversarial inputs through probabilistic decision-path adjustments, TTTS serves as the foundation for FairTTTS. By building on this accuracy-enhancing technique, FairTTTS mitigates bias and improves predictive performance. FairTTTS uses a distance-based heuristic to adjust decisions at protected attribute nodes, ensuring fairness for unprivileged samples. This fairness-oriented adjustment occurs as a post-processing step, allowing FairTTTS to be applied to pre-trained models, diverse datasets, and various fairness metrics without retraining. Extensive evaluation on seven benchmark datasets shows that FairTTTS outperforms traditional methods in fairness improvement, achieving a 20.96% average increase over the baseline compared to 18.78% for related work, and further enhances accuracy by 0.55%. In contrast, competing methods typically reduce accuracy by 0.42%. These results confirm that FairTTTS effectively promotes more equitable decision-making while simultaneously improving predictive performance.

When many people highlight the same document, is the crowd a single consensus, or is it internally structured into reader sub-groups that mark different things -- and is that structure a stable property of a reader or of the document? Building on prior work showing an individual's within-document highlighting signal is a whisper while individuality lives in selection, we ask the group-level question on a co-readership platform using a margin-preserving curveball null. Experiment 1: within a document, readers form strong sub-groups -- pairs agree far beyond what shared salience, mark density, and sentence popularity predict (nearest-neighbour agreement z=+6.3, significant in 88% of documents). Under an eight-block region-preserving null, shared engagement with the same coarse regions of the document accounts for about 40% of this excess; the majority survives as finer reader-specific agreement (z=+3.6, 77% significant). So the within-document crowd is, in a descriptive sense, factional. Experiment 2: is that grouping a stable reader trait? Here we are honest about power. The cross-document split-half reproducibility of a pair's agreement is near zero pooled (+0.078 and 0.000 in two separately drawn samples), and a power calibration shows the test is informative only for pairs that co-read many documents. In the only informative high-overlap subset (k>=4), point estimates are positive but small-sample, imprecise across the separately drawn samples, never significant, and attenuate under the region-preserving null. We therefore leave cross-document stability unresolved: the data is consistent with anything from situational grouping to a weak-to-moderate stable reader trait. The crowd is factional within a document; whether its factions follow the reader across documents is, honestly, beyond our reach.

Overparameterized neural networks can be highly accurate on average on an i.i.d. test set yet consistently fail on atypical groups of the data (e.g., by learning spurious correlations that hold on average but not in such groups). Distributionally robust optimization (DRO) allows us to learn models that instead minimize the worst-case training loss over a set of pre-defined groups. However, we find that naively applying group DRO to overparameterized neural networks fails: these models can perfectly fit the training data, and any model with vanishing average training loss also already has vanishing worst-case training loss. Instead, the poor worst-case performance arises from poor generalization on some groups. By coupling group DRO models with increased regularization---a stronger-than-typical L2 penalty or early stopping---we achieve substantially higher worst-group accuracies, with 10-40 percentage point improvements on a natural language inference task and two image tasks, while maintaining high average accuracies. Our results suggest that regularization is important for worst-group generalization in the overparameterized regime, even if it is not needed for average generalization. Finally, we introduce a stochastic optimization algorithm, with convergence guarantees, to efficiently train group DRO models.

We introduce VideoPrism, a general-purpose video encoder that tackles diverse video understanding tasks with a single frozen model. We pretrain VideoPrism on a heterogeneous corpus containing 36M high-quality video-caption pairs and 582M video clips with noisy parallel text (e.g., ASR transcripts). The pretraining approach improves upon masked autoencoding by global-local distillation of semantic video embeddings and a token shuffling scheme, enabling VideoPrism to focus primarily on the video modality while leveraging the invaluable text associated with videos. We extensively test VideoPrism on four broad groups of video understanding tasks, from web video question answering to CV for science, achieving state-of-the-art performance on 30 out of 33 video understanding benchmarks.

Prompt engineering has emerged as a powerful technique for optimizing large language models (LLMs) for specific applications, enabling faster prototyping and improved performance, and giving rise to the interest of the community in protecting proprietary system prompts. In this work, we explore a novel perspective on prompt privacy through the lens of membership inference. We develop Prompt Detective, a statistical method to reliably determine whether a given system prompt was used by a third-party language model. Our approach relies on a statistical test comparing the distributions of two groups of model outputs corresponding to different system prompts. Through extensive experiments with a variety of language models, we demonstrate the effectiveness of Prompt Detective for prompt membership inference. Our work reveals that even minor changes in system prompts manifest in distinct response distributions, enabling us to verify prompt usage with statistical significance.

On-Policy Self-Distillation (OPSD) has recently emerged as an alternative to Reinforcement Learning with Verifiable Rewards (RLVR), promising higher accuracy and shorter responses through token-level credit assignment from a self-teacher conditioned on privileged context. However, this promise does not carry over to thinking-enabled mathematical reasoning, where reported accuracy gains shrink and sometimes turn negative. We hypothesize that hindsight supervision can specify better token-level alternatives in short thinking-disabled outputs, but in long thinking-enabled traces it more readily identifies redundancy than supplies better replacements. To test this, we applied OPSD separately to correct and incorrect rollout groups, so that compression and correction can be observed in isolation. Our results show that in thinking-enabled mathematical reasoning, OPSD behaves most reliably as a compression mechanism rather than a correction mechanism: training only on correct rollouts preserves accuracy while substantially shortening responses, whereas training only on incorrect rollouts damages accuracy. In light of these findings, we propose a revised post-training pipeline for thinking-enabled mathematical reasoning: SFT then RLVR then OPSD.

Multilingual fluency often invites a stronger assumption: a model that can speak a user's language must also understand the culture encoded by that language. We call this the Illusion of Cultural Alignment. To test this assumption directly, we introduce MSQA, a benchmark of 1,064 natively sourced questions across 11 language groups, five cultural dimensions, and three difficulty tiers. Unlike translated benchmarks, MSQA targets locally grounded knowledge and reduces shortcuts from English-centric cross-lingual transfer. Evaluating 18 LLMs, we find substantial cultural degradation and a pronounced Locality Effect: cultural competence tracks pre-training exposure more closely than general reasoning ability. We further show that common inference-time remedies do not dissolve the illusion. Models remain overconfident on unfamiliar cultural questions, repeated sampling yields unstable rather than reliable correctness, and retrieval augmentation helps unevenly on long-tail facts. These findings indicate that cultural alignment cannot be inferred from multilingual ability alone and requires deeper intervention than calibration, sampling, or retrieval at inference time

Advances in AI-based voice conversion have enabled a wide range of media applications, including films, audiobooks, and games. However, most research and public benchmarks still focus on natural human speech, leaving designed vocalizations, such as monster growls and robotic voices, underexplored, partly due to the lack of publicly available resources. To address this gap, we introduce the Designed Vocalizations Dataset, constructed by curating diverse raw vocal sources, including speech and animal vocalizations, and applying professional vocal effects processing to produce corresponding effect modified variants. We further provide a standardized test set with explicit seen/unseen splits over source timbre groups and preset styles to assess generalization under controlled conditions. Finally, we report baseline benchmark results to support reproducible evaluation and future research. The dataset and demo samples are available at https://ncai-official.github.io/speech/publications/designed-vocalizations-dataset/.

Large language models (LLMs) are increasingly used as judges for subjective tasks, where annotators disagree and the relevant question is not only how accurate a judge is, but whose judgments it reproduces. Sociodemographic prompting conditions the judge on an annotator's demographic profile to align its judgments with the corresponding group's. We test whether this alignment emerges distributionally, comparing the predicted label distributions of 23 open-weight LLMs on three subjective tasks against those of real annotator groups, under three conditions: no demographic information, single-attribute profiles, and intersectional profiles over gender, age, race, and education. Three findings emerge. First, a judge prompted with no demographics is not perspective-neutral: models best reproduce the judgments of White, college-educated annotators. Second, demographic conditioning is asymmetric: it moves the judge toward majority groups and away from minority groups, most strongly on offensiveness, where intersectional profiles amplify the harm. Third, by comparing base and instruct models we identify instruction-tuning as a possible source of the asymmetry. Demographic conditioning should therefore be used with caution to estimate group judgments: conditioning moves predictions away from the reference distributions of the minority groups the method is often invoked to serve.

Text-to-audio-video (T2AV) generation is central to applications such as filmmaking and world modeling. However, current models often fail to produce physically plausible sounds. Previous benchmarks primarily focus on audio-video temporal synchronization, while largely overlooking explicit evaluation of audio-physics grounding, thereby limiting the study of physically plausible audio-visual generation. To address this issue, we present PhyAVBench, the first benchmark that systematically evaluates the audio-physics grounding capabilities of T2AV, image-to-audio-video (I2AV), and video-to-audio (V2A) models. PhyAVBench offers PhyAV-Sound-11K, a new dataset of 25.5 hours of 11,605 audible videos collected from 184 participants to ensure diversity and avoid data leakage. It contains 337 paired-prompt groups with controlled physical variations that drive sound differences, each grounded with an average of 17 videos and spanning 6 audio-physics dimensions and 41 fine-grained test points. Each prompt pair is annotated with the physical factors underlying their acoustic differences. Importantly, PhyAVBench leverages paired text prompts to evaluate this capability. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST) and introduce a novel metric, the Contrastive Physical Response Score (CPRS), which quantifies the acoustic consistency between generated videos and their real-world counterparts. We conduct a comprehensive evaluation of 17 state-of-the-art models. Our results reveal that even leading commercial models struggle with fundamental audio-physical phenomena, exposing a critical gap beyond audio-visual synchronization and pointing to future research directions. We hope PhyAVBench will serve as a foundation for advancing physically grounded audio-visual generation. Prompts, ground-truth, and generated video samples are available at https://github.com/imxtx/PhyAVBench.

Agent self-evolution updates an agent's persistent state from prior experience and reuses it to solve related tasks more effectively. Evaluating self-evolution is difficult: existing benchmarks provide limited coverage of economically valuable task domains, do not always design training and test tasks such that test-time gains can be attributed to training experience, and remain vulnerable to data contamination. We present GDPevo, an evolution-native benchmark grounded in GDP-related enterprise workflows, together with the fully automated data pipeline that generates it. Its core mechanism, rule hybridization, decomposes each enterprise workflow into atomic business rules, distributes subsets of these rules across training tasks, and recombines them in held-out test tasks so that test-time gains are attributable. GDPevo spans CRM, ERP, finance, healthcare, legal, and data-centric workflows. Its V1 release contains 120 tasks in 12 groups, with five training and five held-out test tasks per group. Full automation enables the pipeline to expand the suite to 240 tasks in 24 groups (V2) within two days, providing a practical response to contamination. Using GDPevo, we evaluate four agents, each comprising a harness and a model, under four supervision types. Self-evolution consistently improves held-out accuracy by up to 16.44 percentage points. But the best evolved agents remain far below the fully informed oracle ceiling of 91.6%, indicating that the self-evolution ability of current agents remains far from fully realized. We publicly release the pipeline, benchmark, and full evaluation results at https://github.com/Prism-Shadow/GDPevo.

As machine learning (ML) systems increasingly impact critical sectors such as hiring, financial risk assessments, and criminal justice, the imperative to ensure fairness has intensified due to potential negative implications. While much ML fairness research has focused on enhancing training data and processes, addressing the outputs of already deployed systems has received less attention. This paper introduces 'BiasGuard', a novel approach designed to act as a fairness guardrail in production ML systems. BiasGuard leverages Test-Time Augmentation (TTA) powered by Conditional Generative Adversarial Network (CTGAN), a cutting-edge generative AI model, to synthesize data samples conditioned on inverted protected attribute values, thereby promoting equitable outcomes across diverse groups. This method aims to provide equal opportunities for both privileged and unprivileged groups while significantly enhancing the fairness metrics of deployed systems without the need for retraining. Our comprehensive experimental analysis across diverse datasets reveals that BiasGuard enhances fairness by 31% while only reducing accuracy by 0.09% compared to non-mitigated benchmarks. Additionally, BiasGuard outperforms existing post-processing methods in improving fairness, positioning it as an effective tool to safeguard against biases when retraining the model is impractical.

Numerous datasets have been proposed to combat the spread of online hate. Despite these efforts, a majority of these resources are English-centric, primarily focusing on overt forms of hate. This research gap calls for developing high-quality corpora in diverse languages that also encapsulate more subtle hate expressions. This study introduces K-HATERS, a new corpus for hate speech detection in Korean, comprising approximately 192K news comments with target-specific offensiveness ratings. This resource is the largest offensive language corpus in Korean and is the first to offer target-specific ratings on a three-point Likert scale, enabling the detection of hate expressions in Korean across varying degrees of offensiveness. We conduct experiments showing the effectiveness of the proposed corpus, including a comparison with existing datasets. Additionally, to address potential noise and bias in human annotations, we explore a novel idea of adopting the Cognitive Reflection Test, which is widely used in social science for assessing an individual's cognitive ability, as a proxy of labeling quality. Findings indicate that annotations from individuals with the lowest test scores tend to yield detection models that make biased predictions toward specific target groups and are less accurate. This study contributes to the NLP research on hate speech detection and resource construction. The code and dataset can be accessed at https://github.com/ssu-humane/K-HATERS.

Reinforcement learning (RL) for code agents often uses executable tests to provide binary rewards. With these rewards, Group Relative Policy Optimization (GRPO) assigns identical advantages to test-passing trajectories within each rollout group, overlooking differences in implementation quality and adherence to task requirements. This leaves the policy without a learning signal that favors clean, targeted implementations over those containing unnecessary or out-of-scope changes. We introduce GAGAR, a framework for quality-aware credit redistribution in code agent RL. Built on dynamic sampling that retains groups containing both passing and failing trajectories, GAGAR places all trajectories from each group in a shared workspace, where an SFT-trained agentic grader jointly inspects them and ranks the test-passing candidates. Based on this ranking, we downweight lower-ranked trajectories and proportionally rescale the advantages of all test-passing trajectories to restore their original sum. This sum-preserving redistribution retains the relative weights established by quality-based downweighting while shifting credit toward higher-quality implementations. We evaluate GAGAR at industrial scale using pre-RL SFT checkpoints of MiMo-V2.6-Flash (310B total parameters) and MiMo-V2.6-Pro (1.02T total parameters). Controlled code-only Flash experiments show improved code agent performance, reduced trajectory-length growth, and more stable training. We further apply GAGAR in large-scale mixed-task RL with both Flash and Pro. Our results support combining test-based verification with groupwise agentic grading to improve the quality and stability of code agent RL.

We introduce Triangular Resampling (TR), a post-training method for mitigating long-horizon error accumulation in motion diffusion models. Built on FloodDiffusion's triangular denoising schedule, TR addresses the mismatch between ground-truth-derived training windows and model-generated inference states. Replacing only completed motion history leaves this mismatch unresolved in partially denoised states within the active window. TR therefore extends rollout-based training to these states, using ground-truth clamping to limit excessive drift. For each replayed sample, TR draws one denoising threshold, shared across latent positions and replay updates, and replays multi-step triangular denoising without gradient tracking. After each update, states below the threshold are replaced with noise-matched ground truth, while those at or above it retain model predictions. The resulting latent window enters the standard training update. This rollout construction supports both supervised training (TR) and distribution matching (TR-DMD). On 120-second motion generation from HumanML3D test prompts, TR and TR-DMD achieve state-of-the-art FID AUC within their respective non-DMD and DMD comparison groups. Supervised TR reduces FID AUC by 40.9% and FID degradation slope by 55.3% relative to matched post-training without replay.

Pretrained Language Models (PLMs) harbor inherent social biases that can result in harmful real-world implications. Such social biases are measured through the probability values that PLMs output for different social groups and attributes appearing in a set of test sentences. However, bias testing is currently cumbersome since the test sentences are generated either from a limited set of manual templates or need expensive crowd-sourcing. We instead propose using ChatGPT for the controllable generation of test sentences, given any arbitrary user-specified combination of social groups and attributes appearing in the test sentences. When compared to template-based methods, our approach using ChatGPT for test sentence generation is superior in detecting social bias, especially in challenging settings such as intersectional biases. We present an open-source comprehensive bias testing framework (BiasTestGPT), hosted on HuggingFace, that can be plugged into any open-source PLM for bias testing. User testing with domain experts from various fields has shown their interest in being able to test modern AI for social biases. Our tool has significantly improved their awareness of such biases in PLMs, proving to be learnable and user-friendly. We thus enable seamless open-ended social bias testing of PLMs by domain experts through an automatic large-scale generation of diverse test sentences for any combination of social categories and attributes.

In light of recent breakthroughs in large language models (LLMs) that have revolutionized natural language processing (NLP), there is an urgent need for new benchmarks to keep pace with the fast development of LLMs. In this paper, we propose CFLUE, the Chinese Financial Language Understanding Evaluation benchmark, designed to assess the capability of LLMs across various dimensions. Specifically, CFLUE provides datasets tailored for both knowledge assessment and application assessment. In knowledge assessment, it consists of 38K+ multiple-choice questions with associated solution explanations. These questions serve dual purposes: answer prediction and question reasoning. In application assessment, CFLUE features 16K+ test instances across distinct groups of NLP tasks such as text classification, machine translation, relation extraction, reading comprehension, and text generation. Upon CFLUE, we conduct a thorough evaluation of representative LLMs. The results reveal that only GPT-4 and GPT-4-turbo achieve an accuracy exceeding 60\% in answer prediction for knowledge assessment, suggesting that there is still substantial room for improvement in current LLMs. In application assessment, although GPT-4 and GPT-4-turbo are the top two performers, their considerable advantage over lightweight LLMs is noticeably diminished. The datasets and scripts associated with CFLUE are openly accessible at https://github.com/aliyun/cflue.

Sparse AutoEncoders (SAEs) offer a promising way to inspect language model representations, but it is still unclear what kind of linguistic structure their latents expose. We use part-of-speech (PoS) categories as a controlled test case to study whether morpho-syntactic information is encoded by individual latents or by structured groups of features. We find that PoS distinctions are highly recoverable from SAE activations, but do not align with one-to-one latent / category mappings. This recoverability is not reducible to lexical memorisation, and Open and Closed PoS classes differ substantially. Categories are supported by compact groups of sparse latents, with substantial variation across tags. These groups remain stable on held-out data, while also showing overlap between related categories. Our results show that SAEs localise morpho-syntactic information in a distributed and category-dependent form rather than through atomic grammatical features.

Physical-field reconstruction and forecasting depend on both measurement density and spatial layout, yet evaluation under a single observation pattern does not characterize performance when that pattern changes. We introduce PDE-OBS, an integrated benchmarking platform spanning numerical data generation, model training, and inference and evaluation under varying observation conditions. It combines 560,000 fields and trajectories from seven partial differential equation families with configurable observation operators and seven adapted baseline methods for stationary reconstruction and short-horizon forecasting. Separating observation construction from physical records allows users to specify parameterized patterns and deterministic mixtures for training and testing while preserving prediction targets and data splits. The evaluation protocol uses references trained for each test pattern to compare models on identical test observations and targets, alongside equal-count groups for spatial-layout comparisons. On a 14,000-record subset, we evaluate 441 trained models under nine test patterns, yielding 3,969 evaluations. Mean cross-pattern error exceeds mean matched-pattern error in all 49 PDE-method pairs, and this finding persists in a configuration-matched subset of 117 models. Denser test observations do not consistently reduce error for a fixed model. Mixed-pattern training on five completed pairs reduces large single-pattern transfer errors, although destination-trained references usually remain more accurate. Together, the benchmark and findings support systematic evaluation of observation-pattern sensitivity and provide a reusable workflow for developing methods under changing measurement conditions. Code: https://github.com/ru1ch3n/PDE-OBS.

Vision-language models (VLMs) achieve strong visual reasoning performance, yet subtle changes from routine image capture and processing can alter their reasoning trajectories even when images appear nearly identical. In long-horizon generation, the resulting activation shifts may accumulate across decoding steps, progressively altering reasoning tokens and ultimately changing the final answer, a phenomenon referred to as answer flips. To address this instability, we propose FlipDir (Flip-Direction Steering), a training-free inference-time method that estimates a low-rank flip-inducing activation subspace from contrastive pairs of original and answer-flipping inputs and selectively steers hidden states during decoding. A margin-based gate limits subspace attenuation to uncertain decoding steps, recovering original predictions while preserving stable ones. To evaluate robustness beyond accuracy or consistency on fixed test sets, we introduce VisFlip, a benchmark framework that constructs evaluation groups for a target model and visual variation setting to separately assess recovery of original predictions and preservation of stable ones. VisFlip spans nine dataset-variation combinations across scientific reasoning, robot-scene understanding, and medical VQA, covering subtle visual variations common in each domain. Experiments across 18 settings demonstrate that FlipDir consistently outperforms existing methods on the combined recovery and preservation metric. We will make our code publicly available.

Language Models (LMs) often cannot be deployed because of their potential to harm users in hard-to-predict ways. Prior work identifies harmful behaviors before deployment by using human annotators to hand-write test cases. However, human annotation is expensive, limiting the number and diversity of test cases. In this work, we automatically find cases where a target LM behaves in a harmful way, by generating test cases ("red teaming") using another LM. We evaluate the target LM's replies to generated test questions using a classifier trained to detect offensive content, uncovering tens of thousands of offensive replies in a 280B parameter LM chatbot. We explore several methods, from zero-shot generation to reinforcement learning, for generating test cases with varying levels of diversity and difficulty. Furthermore, we use prompt engineering to control LM-generated test cases to uncover a variety of other harms, automatically finding groups of people that the chatbot discusses in offensive ways, personal and hospital phone numbers generated as the chatbot's own contact info, leakage of private training data in generated text, and harms that occur over the course of a conversation. Overall, LM-based red teaming is one promising tool (among many needed) for finding and fixing diverse, undesirable LM behaviors before impacting users.

Learning from Label Proportions (LLP) is a learning problem where only aggregate level labels are available for groups of instances, called bags, during training, and the aim is to get the best performance at the instance-level on the test data. This setting arises in domains like advertising and medicine due to privacy considerations. We propose a novel algorithmic framework for this problem that iteratively performs two main steps. For the first step (Pseudo Labeling) in every iteration, we define a Gibbs distribution over binary instance labels that incorporates a) covariate information through the constraint that instances with similar covariates should have similar labels and b) the bag level aggregated label. We then use Belief Propagation (BP) to marginalize the Gibbs distribution to obtain pseudo labels. In the second step (Embedding Refinement), we use the pseudo labels to provide supervision for a learner that yields a better embedding. Further, we iterate on the two steps again by using the second step's embeddings as new covariates for the next iteration. In the final iteration, a classifier is trained using the pseudo labels. Our algorithm displays strong gains against several SOTA baselines (up to 15%) for the LLP Binary Classification problem on various dataset types - tabular and Image. We achieve these improvements with minimal computational overhead above standard supervised learning due to Belief Propagation, for large bag sizes, even for a million samples.

A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment. The gap is especially important for AI agents, whose observations, tool outputs, documents, and prior decisions accumulate over long trajectories. LongStraw is an architecture-aware execution stack for million-token RL post-training under a fixed GPU budget, instantiated with Group Relative Policy Optimization (GRPO). It evaluates the shared prompt without autograd, retains only model-specific state needed by later tokens, and replays short response branches one at a time, reducing the live training graph at the cost of additional replay time. We implement it for the hybrid recurrent and full-attention Qwen3.6-27B and the compressed-attention mixture-of-experts GLM-5.2. On eight H20 GPUs, LongStraw completes grouped Qwen scoring and response backward at 2.1M positions for groups of 2 and 8; increasing the group size adds only 0.21 GB of peak allocated memory, while a separate stress test reaches 4.46M positions. On 32 H20 GPUs, we validate the end-to-end LongStraw execution path for a 2.1M-token prompt across all 78 layers of GLM-5.2. These experiments establish execution capacity rather than complete training correctness because the captured prompt state is detached and some distributed forward and gradient composition paths remain incomplete.

Although media bias detection is a complex multi-task problem, there is, to date, no unified benchmark grouping these evaluation tasks. We introduce the Media Bias Identification Benchmark (MBIB), a comprehensive benchmark that groups different types of media bias (e.g., linguistic, cognitive, political) under a common framework to test how prospective detection techniques generalize. After reviewing 115 datasets, we select nine tasks and carefully propose 22 associated datasets for evaluating media bias detection techniques. We evaluate MBIB using state-of-the-art Transformer techniques (e.g., T5, BART). Our results suggest that while hate speech, racial bias, and gender bias are easier to detect, models struggle to handle certain bias types, e.g., cognitive and political bias. However, our results show that no single technique can outperform all the others significantly. We also find an uneven distribution of research interest and resource allocation to the individual tasks in media bias. A unified benchmark encourages the development of more robust systems and shifts the current paradigm in media bias detection evaluation towards solutions that tackle not one but multiple media bias types simultaneously.

In this work, we propose an approach for collecting completion usage logs from the users in an IDE and using them to train a machine learning based model for ranking completion candidates. We developed a set of features that describe completion candidates and their context, and deployed their anonymized collection in the Early Access Program of IntelliJ-based IDEs. We used the logs to collect a dataset of code completions from users, and employed it to train a ranking CatBoost model. Then, we evaluated it in two settings: on a held-out set of the collected completions and in a separate A/B test on two different groups of users in the IDE. Our evaluation shows that using a simple ranking model trained on the past user behavior logs significantly improved code completion experience. Compared to the default heuristics-based ranking, our model demonstrated a decrease in the number of typing actions necessary to perform the completion in the IDE from 2.073 to 1.832. The approach adheres to privacy requirements and legal constraints, since it does not require collecting personal information, performing all the necessary anonymization on the client's side. Importantly, it can be improved continuously: implementing new features, collecting new data, and evaluating new models - this way, we have been using it in production since the end of 2020.

Synthetic data generation has the potential to impact applications and domains with scarce data. However, before such data is used for sensitive tasks such as mental health, we need an understanding of how different demographics are represented in it. In our paper, we analyze the potential of producing synthetic data using GPT-3 by exploring the various stressors it attributes to different race and gender combinations, to provide insight for future researchers looking into using LLMs for data generation. Using GPT-3, we develop HEADROOM, a synthetic dataset of 3,120 posts about depression-triggering stressors, by controlling for race, gender, and time frame (before and after COVID-19). Using this dataset, we conduct semantic and lexical analyses to (1) identify the predominant stressors for each demographic group; and (2) compare our synthetic data to a human-generated dataset. We present the procedures to generate queries to develop depression data using GPT-3, and conduct analyzes to uncover the types of stressors it assigns to demographic groups, which could be used to test the limitations of LLMs for synthetic data generation for depression data. Our findings show that synthetic data mimics some of the human-generated data distribution for the predominant depression stressors across diverse demographics.

Dense per-step supervision is an appealing remedy for sparse-reward, long-horizon LLM agents: reward the agent for predicting its next observation, and memory should follow. We show that under group-normalized RL (GRPO), this recipe does not merely fail -- it destroys the policy. Across Qwen3-1.7B/4B/8B on ALFWorld, a potential-based prediction reward drives every run into a degenerate absorbing state (prediction accuracy -> 1.0, task success -> 0,episode length pinned at the horizon): the "dark room" pathology, built automatically by the optimizer. A single-factor ablation localizes the cause -- removing only GRPO's std normalization turns the same reward from catastrophic (0%) into baseline parity -- and a two-line proposition explains why: in all-fail groups the z-scored advantage is invariant to the shaping coefficient, so bounded rewards become unbounded pressure and annealing cannot help. Our central insight generalizes this: what z-scoring amplifies is a dense signal's within-group variance while all-fail groups dominate, so signals whose variance decays by mastery are structurally amplifier-safe.This variance-profile criterion retrodicts our collapses, carries preregistered predictions for arms that had not yet run, and is consistent with published reward-channel successes (a compatibility check, not an independent test). Finally, a controlled signal-delivery matrix (identical signal, varying only the consumption mechanism) shows the reward channel is at best neutral while the auxiliary-loss channel gains ~20 points -- and a shuffled-gold placebo matches the true-gold arm, so the gap survives without correct labels. Endpoints are single-seed; seed replication and group-size controls are preregistered and in progress.

The robustness of machine learning models can be compromised by spurious correlations between non-causal features in the input data and target labels. A common way to test for such correlations is to train on data where the label is strongly tied to some non-causal cue, then evaluate on examples where that tie no longer holds. This idea is well established for classification tasks, but for semantic segmentation the specific failure modes are not well understood. We show that a model may achieve reasonable overlap while assigning the wrong semantic label, swapping one plausible foreground class for another, even when object boundaries are largely correct. We focus on this semantic label-flip behaviour and quantify it with a simple diagnostic (Flip) that counts how often ground truth foreground pixels are assigned the wrong foreground identity while remaining predicted as foreground. In a setting where category and scene are correlated during training, increasing the correlation consistently widens the gap between common and rare test conditions and increases these within-object label swaps on counterfactual groups. Overall, our results motivate assessing segmentation robustness under distribution shift beyond overlap by decomposing foreground errors into correct pixels, flipped-identity pixels, and missed-to-background pixels. We also propose an entropy-based, ground truth label-free `flip-risk' score, which is computed from foreground identity uncertainty, and show that it can flag flip-prone cases at inference time. Code is available at https://github.com/acharaakshit/label-flips.

Large language model (LLM) providers boast big numbers for maximum context window sizes. To test the real world use of context windows, we 1) define a concept of maximum effective context window, 2) formulate a testing method of a context window's effectiveness over various sizes and problem types, and 3) create a standardized way to compare model efficacy for increasingly larger context window sizes to find the point of failure. We collected hundreds of thousands of data points across several models and found significant differences between reported Maximum Context Window (MCW) size and Maximum Effective Context Window (MECW) size. Our findings show that the MECW is, not only, drastically different from the MCW but also shifts based on the problem type. A few top of the line models in our test group failed with as little as 100 tokens in context; most had severe degradation in accuracy by 1000 tokens in context. All models fell far short of their Maximum Context Window by as much as 99 percent. Our data reveals the Maximum Effective Context Window shifts based on the type of problem provided, offering clear and actionable insights into how to improve model accuracy and decrease model hallucination rates.