Daily Papers

Conventional agent systems often struggle in open-ended environments where task distributions continuously drift and external supervision is scarce. Their reliance on static toolsets or offline training lags behind these dynamics, leaving the system's capability boundaries rigid and unknown. To address this, we propose the In-Situ Self-Evolving paradigm. This approach treats sequential task interactions as a continuous stream of experience, enabling the system to distill short-term execution feedback into long-term, reusable capabilities without access to ground-truth labels. Within this framework, we identify tool evolution as the critical pathway for capability expansion, which provides verifiable, binary feedback signals. Within this framework, we develop Yunjue Agent, a system that iteratively synthesizes, optimizes, and reuses tools to navigate emerging challenges. To optimize evolutionary efficiency, we further introduce a Parallel Batch Evolution strategy. Empirical evaluations across five diverse benchmarks under a zero-start setting demonstrate significant performance gains over proprietary baselines. Additionally, complementary warm-start evaluations confirm that the accumulated general knowledge can be seamlessly transferred to novel domains. Finally, we propose a novel metric to monitor evolution convergence, serving as a function analogous to training loss in conventional optimization. We open-source our codebase, system traces, and evolved tools to facilitate future research in resilient, self-evolving intelligence.

We train several language models for Icelandic, including IceBERT, that achieve state-of-the-art performance in a variety of downstream tasks, including part-of-speech tagging, named entity recognition, grammatical error detection and constituency parsing. To train the models we introduce a new corpus of Icelandic text, the Icelandic Common Crawl Corpus (IC3), a collection of high quality texts found online by targeting the Icelandic top-level-___domain (TLD). Several other public data sources are also collected for a total of 16GB of Icelandic text. To enhance the evaluation of model performance and to raise the bar in baselines for Icelandic, we translate and adapt the WinoGrande dataset for co-reference resolution. Through these efforts we demonstrate that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus. We further show that initializing models using existing multilingual models can lead to state-of-the-art results for some downstream tasks.

We present a faster algorithm to generate a warm start for sampling an arbitrary logconcave density specified by an evaluation oracle, leading to the first sub-cubic sampling algorithms for inputs in (near-)isotropic position. A long line of prior work incurred a warm-start penalty of at least linear in the dimension, hitting a cubic barrier, even for the special case of uniform sampling from convex bodies. Our improvement relies on two key ingredients of independent interest. (1) We show how to sample given a warm start in weaker notions of distance, in particular q-R\'enyi divergence for q=mathcal{O}(1), whereas previous analyses required stringent infty-R\'enyi divergence (with the exception of Hit-and-Run, whose known mixing time is higher). This marks the first improvement in the required warmness since Lov\'asz and Simonovits (1991). (2) We refine and generalize the log-Sobolev inequality of Lee and Vempala (2018), originally established for isotropic logconcave distributions in terms of the diameter of the support, to logconcave distributions in terms of a geometric average of the support diameter and the largest eigenvalue of the covariance matrix.

Methods for making high-quality recommendations often rely on learning latent representations from interaction data. These methods, while performant, do not provide ready mechanisms for users to control the recommendation they receive. Our work tackles this problem by proposing LACE, a novel concept value bottleneck model for controllable text recommendations. LACE represents each user with a succinct set of human-readable concepts through retrieval given user-interacted documents and learns personalized representations of the concepts based on user documents. This concept based user profile is then leveraged to make recommendations. The design of our model affords control over the recommendations through a number of intuitive interactions with a transparent user profile. We first establish the quality of recommendations obtained from LACE in an offline evaluation on three recommendation tasks spanning six datasets in warm-start, cold-start, and zero-shot setups. Next, we validate the controllability of LACE under simulated user interactions. Finally, we implement LACE in an interactive controllable recommender system and conduct a user study to demonstrate that users are able to improve the quality of recommendations they receive through interactions with an editable user profile.

In many real-world deployments of machine learning systems, data arrive piecemeal. These learning scenarios may be passive, where data arrive incrementally due to structural properties of the problem (e.g., daily financial data) or active, where samples are selected according to a measure of their quality (e.g., experimental design). In both of these cases, we are building a sequence of models that incorporate an increasing amount of data. We would like each of these models in the sequence to be performant and take advantage of all the data that are available to that point. Conventional intuition suggests that when solving a sequence of related optimization problems of this form, it should be possible to initialize using the solution of the previous iterate -- to "warm start" the optimization rather than initialize from scratch -- and see reductions in wall-clock time. However, in practice this warm-starting seems to yield poorer generalization performance than models that have fresh random initializations, even though the final training losses are similar. While it appears that some hyperparameter settings allow a practitioner to close this generalization gap, they seem to only do so in regimes that damage the wall-clock gains of the warm start. Nevertheless, it is highly desirable to be able to warm-start neural network training, as it would dramatically reduce the resource usage associated with the construction of performant deep learning systems. In this work, we take a closer look at this empirical phenomenon and try to understand when and how it occurs. We also provide a surprisingly simple trick that overcomes this pathology in several important situations, and present experiments that elucidate some of its properties.

Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial warm restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple warm restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on the CIFAR-10 and CIFAR-100 datasets, where we demonstrate new state-of-the-art results at 3.14% and 16.21%, respectively. We also demonstrate its advantages on a dataset of EEG recordings and on a downsampled version of the ImageNet dataset. Our source code is available at https://github.com/loshchil/SGDR

Large language models (LLMs) are routinely pre-trained on billions of tokens, only to restart the process over again once new data becomes available. A much cheaper and more efficient solution would be to enable the continual pre-training of these models, i.e. updating pre-trained models with new data instead of re-training them from scratch. However, the distribution shift induced by novel data typically results in degraded performance on past data. Taking a step towards efficient continual pre-training, in this work, we examine the effect of different warm-up strategies. Our hypothesis is that the learning rate must be re-increased to improve compute efficiency when training on a new dataset. We study the warmup phase of models pre-trained on the Pile (upstream data, 300B tokens) as we continue to pre-train on SlimPajama (downstream data, 297B tokens), following a linear warmup and cosine decay schedule. We conduct all experiments on the Pythia 410M language model architecture and evaluate performance through validation perplexity. We experiment with different pre-training checkpoints, various maximum learning rates, and various warmup lengths. Our results show that while rewarming models first increases the loss on upstream and downstream data, in the longer run it improves the downstream performance, outperforming models trained from scratchx2013even for a large downstream dataset.

Adapting reasoning models to new tasks during post-training with only output-level supervision stalls under reinforcement learning from verifiable rewards (RLVR) when the initial success probability p_0 is small. Using the Tsallis q-logarithm, we define a loss family J_Q that interpolates between RLVR (at q{=}0, the exploitation pole) and the log-marginal-likelihood over latent trajectories (at q{=}1, the density-estimation pole). All members share the same per-example gradient direction, differing only by a scalar amplification P_{θ^{-q}} that reweights each instance independently of the learning rate. This amplification is the mechanism that addresses cold-start stalling: under gradient flow, the exploitation pole requires Ω(1{p_0}) time to escape cold start, while the density-estimation pole escapes in Θbig(log(1{p_0})big); intermediate q trades escape speed against noise memorization. Because P_θ is intractable, we derive two Monte Carlo estimators from the two factorizations of the gradient: Gradient-Amplified RL (GARL) samples from the prior and amplifies the RL gradient, and Posterior-Attenuated Fine-Tuning (PAFT) importance-resamples from the posterior and runs standard SFT. Both have bias Obig(q{M P_θ^{q+1}}big); GARL has lower variance, PAFT has semantically coherent gradients. On FinQA, HotPotQA, and MuSiQue, GARL at q{=}0.75 substantially mitigates cold-start stalling, escaping cold start where GRPO fails entirely. In warm start, GARL at low q dominates FinQA where training is stable; on HotPotQA and MuSiQue, GARL destabilizes during training, and PAFT at q{=}0.75 provides stable gradients (best overall on HotPotQA at 47.9 maj@16, +14.4 over GRPO).

Small language models are attractive for production deployment due to their low cost, fast inference, and ease of specialization. However, adapting them to a specific task remains a challenging engineering loop, driven not by training itself but by surrounding decisions: data curation, failure diagnosis, regression avoidance, and iteration control. We present Pioneer Agent, a closed-loop system that automates this lifecycle. In cold-start mode, given only a natural-language task description, the agent acquires data, constructs evaluation sets, and iteratively trains models by jointly optimizing data, hyperparameters, and learning strategy. In production mode, given a deployed model with labeled failures, it diagnoses error patterns, constructs targeted training data, and retrains under explicit regression constraints. To evaluate this setting, we introduce AdaptFT-Bench, a benchmark of synthetic inference logs with progressively increasing noise, designed to test the full adaptation loop: diagnosis, curriculum synthesis, retraining, and verification. Across eight cold-start benchmarks spanning reasoning, math, code generation, summarization, and classification, Pioneer Agent improves over base models by 1.6-83.8 points. On AdaptFT-Bench, it improves or preserves performance in all seven scenarios, while naive retraining degrades by up to 43 points. On two production-style deployments built from public benchmark tasks, it raises intent classification from 84.9% to 99.3% and Entity F1 from 0.345 to 0.810. Beyond performance gains, the agent often discovers effective training strategies, including chain-of-thought supervision, task-specific optimization, and quality-focused data curation, purely from downstream feedback.

The learning rate warmup heuristic achieves remarkable success in stabilizing training, accelerating convergence and improving generalization for adaptive stochastic optimization algorithms like RMSprop and Adam. Here, we study its mechanism in details. Pursuing the theory behind warmup, we identify a problem of the adaptive learning rate (i.e., it has problematically large variance in the early stage), suggest warmup works as a variance reduction technique, and provide both empirical and theoretical evidence to verify our hypothesis. We further propose RAdam, a new variant of Adam, by introducing a term to rectify the variance of the adaptive learning rate. Extensive experimental results on image classification, language modeling, and neural machine translation verify our intuition and demonstrate the effectiveness and robustness of our proposed method. All implementations are available at: https://github.com/LiyuanLucasLiu/RAdam.

Time intervals between purchasing items are a crucial factor in sequential recommendation tasks, whereas existing approaches focus on item sequences and often overlook by assuming the intervals between items are static. However, dynamic intervals serve as a dimension that describes user profiling on not only the history within a user but also different users with the same item history. In this work, we propose IntervalLLM, a novel framework that integrates interval information into LLM and incorporates the novel interval-infused attention to jointly consider information of items and intervals. Furthermore, unlike prior studies that address the cold-start scenario only from the perspectives of users and items, we introduce a new viewpoint: the interval perspective to serve as an additional metric for evaluating recommendation methods on the warm and cold scenarios. Extensive experiments on 3 benchmarks with both traditional- and LLM-based baselines demonstrate that our IntervalLLM achieves not only 4.4% improvements in average but also the best-performing warm and cold scenarios across all users, items, and the proposed interval perspectives. In addition, we observe that the cold scenario from the interval perspective experiences the most significant performance drop among all recommendation methods. This finding underscores the necessity of further research on interval-based cold challenges and our integration of interval information in the realm of sequential recommendation tasks. Our code is available here: https://github.com/sony/ds-research-code/tree/master/recsys25-IntervalLLM.

Climate-Eval is a comprehensive benchmark designed to evaluate natural language processing models across a broad range of tasks related to climate change. Climate-Eval aggregates existing datasets along with a newly developed news classification dataset, created specifically for this release. This results in a benchmark of 25 tasks based on 13 datasets, covering key aspects of climate discourse, including text classification, question answering, and information extraction. Our benchmark provides a standardized evaluation suite for systematically assessing the performance of large language models (LLMs) on these tasks. Additionally, we conduct an extensive evaluation of open-source LLMs (ranging from 2B to 70B parameters) in both zero-shot and few-shot settings, analyzing their strengths and limitations in the ___domain of climate change.

Benchmarks are necessary for healthcare evaluation, but are not sufficient for predicting deployment performance. Our position is that the evaluation--deployment gap arises not because of poorly designed benchmarks, but from implicit assumptions about how users interact with models that cannot be surfaced from benchmarks alone. To make this precise, we propose a classification of assumptions into two categories: task, which can be tested from conversation data alone, and outcome, which requires outcome data and behavioral studies for testing. Critically, outcome assumptions depend on human behavior, something that even well-designed benchmarks cannot directly observe. To demonstrate the operationality of this framework, we retrospectively analyze a healthcare RCT as a case study and find that the gap naturally separates into task and outcome gaps of roughly equal size. To address this, we make two contributions: first, we propose BenchmarkCards, an artifact that documents assumptions, and second, we propose staged evaluation, a procedure that systematically tests assumptions and evaluates performance.

The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the ___location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications.

The prohibitive cost of evaluating large language models (LLMs) on comprehensive benchmarks necessitates the creation of small yet representative data subsets (i.e., tiny benchmarks) that enable efficient assessment while retaining predictive fidelity. Current methods for this task operate under a model-centric paradigm, selecting benchmarking items based on the collective performance of existing models. Such approaches are limited by large upfront costs, an inability to immediately handle new benchmarks (`cold-start'), and the fragile assumption that future models will share the failure patterns of their predecessors. In this work, we challenge this paradigm and propose a item-centric approach to benchmark subset selection, arguing that selection should be based on the intrinsic properties of the task items themselves, rather than on model-specific failure patterns. We instantiate this item-centric efficient benchmarking approach via a novel method, Scales++, where data selection is based on the cognitive demands of the benchmark samples. Empirically, we show Scales++ reduces the upfront selection cost by over 18x while achieving competitive predictive fidelity. On the Open LLM Leaderboard, using just a 0.5\% data subset, we predict full benchmark scores with a 2.9% mean absolute error. We demonstrate that this item-centric approach enables more efficient model evaluation without significant fidelity degradation, while also providing better cold-start performance and more interpretable benchmarking.

Evaluations are critical for understanding the capabilities of large language models (LLMs). Fundamentally, evaluations are experiments; but the literature on evaluations has largely ignored the literature from other sciences on experiment analysis and planning. This article shows researchers with some training in statistics how to think about and analyze data from language model evaluations. Conceptualizing evaluation questions as having been drawn from an unseen super-population, we present formulas for analyzing evaluation data, measuring differences between two models, and planning an evaluation experiment. We make a number of specific recommendations for running language model evaluations and reporting experiment results in a way that minimizes statistical noise and maximizes informativeness.

The ``LLM-as-a-judge'' paradigm has become a standard method for evaluating open-ended generation. To address the quadratic scalability costs of pairwise comparisons, popular benchmarks like Arena-Hard and AlpacaEval compare all models against a single anchor. However, despite its widespread use, the impact of anchor selection on the reliability of the results remains largely unexplored. In this work, we systematically investigate the effect of anchor selection by evaluating 22 different anchors on the Arena-Hard-v2.0 dataset. We find that the choice of anchor is critical: a poor anchor can dramatically reduce correlation with human rankings. We identify that common anchor choices (best-performing and worst-performing models) make poor anchors. Because these extreme anchors are consistently better or worse than all other models, they are seldom indicative of the relative ranking of the models. We further quantify the effect size of anchor selection, showing it is comparable to the selection of a judge model. We conclude with actionable recommendations. First, we conduct a power analysis, and compute sufficient benchmark sizes for anchor-based evaluation, finding that standard benchmark sizes are insufficient for pairwise evaluation and fail to distinguish between competitive models reliably. Second, we provide guidelines for selecting informative anchors to ensure reliable and efficient evaluation practices.

In this work, we study the evolution of the loss Hessian across many classification tasks in order to understand the effect the curvature of the loss has on the training dynamics. Whereas prior work has focused on how different learning rates affect the loss Hessian observed during training, we also analyze the effects of model initialization, architectural choices, and common training heuristics such as gradient clipping and learning rate warmup. Our results demonstrate that successful model and hyperparameter choices allow the early optimization trajectory to either avoid -- or navigate out of -- regions of high curvature and into flatter regions that tolerate a higher learning rate. Our results suggest a unifying perspective on how disparate mitigation strategies for training instability ultimately address the same underlying failure mode of neural network optimization, namely poor conditioning. Inspired by the conditioning perspective, we show that learning rate warmup can improve training stability just as much as batch normalization, layer normalization, MetaInit, GradInit, and Fixup initialization.

Artificial intelligence (AI) developers are increasingly building language models with warm and empathetic personas that millions of people now use for advice, therapy, and companionship. Here, we show how this creates a significant trade-off: optimizing language models for warmth undermines their reliability, especially when users express vulnerability. We conducted controlled experiments on five language models of varying sizes and architectures, training them to produce warmer, more empathetic responses, then evaluating them on safety-critical tasks. Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing incorrect factual information, and offering problematic medical advice. They were also significantly more likely to validate incorrect user beliefs, particularly when user messages expressed sadness. Importantly, these effects were consistent across different model architectures, and occurred despite preserved performance on standard benchmarks, revealing systematic risks that current evaluation practices may fail to detect. As human-like AI systems are deployed at an unprecedented scale, our findings indicate a need to rethink how we develop and oversee these systems that are reshaping human relationships and social interaction.

Recent machine learning papers often report 1-2 percentage point improvements from a single run on a benchmark. These gains are highly sensitive to random seeds, data ordering, and implementation details, yet are rarely accompanied by uncertainty estimates or significance tests. It is therefore unclear when a reported +1-2% reflects a real algorithmic advance versus noise. We revisit this problem under realistic compute budgets, where only a few runs are affordable. We propose a simple, PC-friendly evaluation protocol based on paired multi-seed runs, bias-corrected and accelerated (BCa) bootstrap confidence intervals, and a sign-flip permutation test on per-seed deltas. The protocol is intentionally conservative and is meant as a guardrail against over-claiming. We instantiate it on CIFAR-10, CIFAR-10N, and AG News using synthetic no-improvement, small-gain, and medium-gain scenarios. Single runs and unpaired t-tests often suggest significant gains for 0.6-2.0 point improvements, especially on text. With only three seeds, our paired protocol never declares significance in these settings. We argue that such conservative evaluation is a safer default for small gains under tight budgets.

Time series foundation models excel in zero-shot forecasting, handling diverse tasks without explicit training. However, the advancement of these models has been hindered by the lack of comprehensive benchmarks. To address this gap, we introduce the General Time Series Forecasting Model Evaluation, GIFT-Eval, a pioneering benchmark aimed at promoting evaluation across diverse datasets. GIFT-Eval encompasses 28 datasets over 144,000 time series and 177 million data points, spanning seven domains, 10 frequencies, multivariate inputs, and prediction lengths ranging from short to long-term forecasts. To facilitate the effective pretraining and evaluation of foundation models, we also provide a non-leaking pretraining dataset containing approximately 230 billion data points. Additionally, we provide a comprehensive analysis of 17 baselines, which includes statistical models, deep learning models, and foundation models. We discuss each model in the context of various benchmark characteristics and offer a qualitative analysis that spans both deep learning and foundation models. We believe the insights from this analysis, along with access to this new standard zero-shot time series forecasting benchmark, will guide future developments in time series foundation models. The codebase, datasets, and a leaderboard showing all the results in detail will be available soon.

We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of practices that utilize knowledge about evaluation tasks at training time. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data prior to evaluation. We then show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models, with broad implications for benchmarking and the study of emergent capabilities.

Agentic systems are evaluated on benchmarks where agents interact with environments to solve tasks. Most papers report a pass@1 score computed from a single run per task, assuming this gives a reliable performance estimate. We test this assumption by collecting 60,000 agentic trajectories on SWE-Bench-Verified, spanning three models and two scaffolds. We find substantial variance: single-run pass@1 estimates vary by 2.2 to 6.0 percentage points depending on which run is selected, with standard deviations exceeding 1.5 percentage points even at temperature 0. This variance has critical implications: reported improvements of 2--3 percentage points may reflect evaluation noise rather than genuine algorithmic progress. Through token-level analysis, we show that trajectories diverge early, often within the first few percent of tokens, and that these small differences cascade into different solution strategies. To enable reliable evaluation of agentic systems, we recommend three concrete practices: (1) estimate pass@1 from multiple independent runs per task, especially when measuring small improvements, (2) use statistical power analysis to determine the number of runs needed to detect expected effect sizes, and (3) consider metrics like pass@k (optimistic bound) and pass^k (pessimistic bound) with k>1 to better characterize the full performance envelope. While these practices increase evaluation cost, they are essential for distinguishing genuine scientific progress from statistical noise.

Forecasting startup success is notoriously difficult, partly because meaningful outcomes, such as exits, large funding rounds, and sustained revenue growth, are rare and can take years to materialize. As a result, signals are sparse and evaluation cycles are slow. Y Combinator batches offer a unique mitigation: each batch comprises around 200 startups, funded simultaneously, with evaluation at Demo Day only three months later. We introduce YC Bench, a live benchmark for forecasting early outperformance within YC batches. Using the YC W26 batch as a case study (196 startups), we measure outperformance with a Pre-Demo Day Score, a KPI combining publicly available traction signals and web visibility. This short-term metric enables rapid evaluation of forecasting models. As a baseline, we take Google mentions prior to the YC W26 application deadline, a simple proxy for prior brand recognition, recovering 6 of 11 top performers at YC Demo Day (55% recall). YC Bench provides a live benchmark for studying startup success forecasting, with iteration cycles measured in months rather than years. Code and Data are available on GitHub: https://github.com/benstaf/ycbench

Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM's activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on documents with factual descriptions of the model (1) using Python type hints during evaluation but not during deployment and (2) recognizing that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. We find that activation steering can suppress evaluation awareness and make the model act like it is deployed even when the cue is present. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

The large language model (LLM)-as-judge paradigm has been used to meet the demand for a cheap, reliable, and fast evaluation of model outputs during AI system development and post-deployment monitoring. While judge models -- LLMs finetuned to specialize in assessing and critiquing model outputs -- have been touted as general purpose evaluators, they are typically evaluated only on non-contextual scenarios, such as instruction following. The omission of contextual settings -- those where external information is used as context to generate an output -- is surprising given the increasing prevalence of retrieval-augmented generation (RAG) and summarization use cases. Contextual assessment is uniquely challenging, as evaluation often depends on practitioner priorities, leading to conditional evaluation criteria (e.g., comparing responses based on factuality and then considering completeness if they are equally factual). To address the gap, we propose ContextualJudgeBench, a judge benchmark with 2,000 challenging response pairs across eight splits inspired by real-world contextual evaluation scenarios. We build our benchmark with a multi-pronged data construction pipeline that leverages both existing human annotations and model-based perturbations. Our comprehensive study across 11 judge models and 9 general purpose models, reveals that the contextual information and its assessment criteria present a significant challenge to even state-of-the-art models. For example, OpenAI's o1, the best-performing model, barely reaches 55% consistent accuracy.