arXiv subjects

Andrew Rabinovich

Publications and source records attributed to Andrew Rabinovich.

At least 19 recordsLinked to original sources

arXiv2026

Fragility Is Not Reducible to Hardness: Conditional Causal Difficulty for LLMs

Large language models show large and seemingly arbitrary performance drops when a task is rewritten in ways that leave its requirements untouched. We recast task difficulty as a causal quantity: the treatment effect of meaning-preserving perturbations on a model. A meaning-preserving rewrite can only be observed to hurt a task the model can already solve, so we place the estimand on the principal stratum where the degradation occurs: the conditional causal difficulty (CCD), the probability that a randomly drawn twin breaks a solvable task. Because harder tasks tend to be more phrasing-sensitive, complexity enters as an effect modifier rather than a confounder, and we residualize it out to obtain complexity-residualized difficulty (CRD); zero for a task exactly as fragile as its complexity predicts, positive for anomalously fragile ones. Current benchmarks do not supply the data CCD requires, so we introduce GlassGSM and GlassMATH, augmenting GSM8K and MATH with twins built from perturbation families, raising fragility on GlassGSM by $\approx 12\times$ over GSM-IC-style distractors on the same problems ($3.7\%\to44.9\%$) while keeping the answer key unchanged. Our predictor Frail learns the per-perturbation break probability and reaches $0.613{\pm}0.023$ of the suite reliability ceiling on GlassGSM and $0.557{\pm}0.023$ on GlassMATH, against $0.171$/$0.401$ for a surface-feature baseline and $0.444$/$0.463$ for supervised baselines such as VAP3, and, at a $5\%$ curation budget on GlassGSM, doubles random recall of fragile items, matching an oracle ranking. Additionally, cross-fitted residualization leaves CRD uncorrelated with the declared proxy in practice. Code is available at https://github.com/upwork/Conditional_Causal_Difficulty_for_LLMs

cs.LG↗

arXiv2026

Quantiles, Not Points: Streaming Quantile Matching for Ordinal-Aware LLM-as-a-Judge Calibration

LLM-as-a-judge scores are often distributionally shifted from human ratings on any given evaluation task, and post-hoc calibration (mapping LLM to human scores in a small labeled anchor set) has become a popular method to correct for this shift. In this paper we derive the calibration from two positions: that LLMaJ scores are ordinal ratings drawn from a distribution rather than isolated points. We present streaming quantile matching (SQM), a version of a classical method not previously bench marked for LLMaJ calibration, which is built on the two positions by computing the optimal-transport map between the LLM and human score distributions. We develop an evaluation framework and within it we show that unlike point-wise calibrations, SQM respects the ordinality of the human anchor set and achieves the best distributional fidelity and the top overall rank across models and data. We also show that in general, post-hoc calibration narrows the raw error gap between smaller and larger judge models, which opens the possibility of more efficient LLM judges.

cs.AI↗

arXiv2026

Prompt Lottery in LLM-as-a-Judge: Measuring and Mitigating Sensitivity to Syntactic Rephrasing

LLM-as-a-Judge has become a standard for evaluating large language models, yet most approaches implicitly assume that a single judge prompt functions as a stable evaluation instrument. Prior work has primarily explored deliberate prompt modifications such as scoring rubrics, role descriptions, or chain-of-thought reasoning in order to optimize accuracy. In contrast, we investigate whether validated prompt variants intended to preserve evaluation semantics while differing in surface realization (\textit{paraphrastic rephrasings}) yield consistent judgments under a fixed evaluation task. We introduce \textit{Prompt Lottery}, a phenomenon in which paraphrastic rephrasings produce different LLMaJ judgments. Across a large-scale study spanning diverse evaluation tasks, judge families, and model backbones, we find that these variations can alter evaluation outcomes and, in some settings, reverse relative model rankings. A user study with 10 human annotators shows that LLMaJ judges exhibit 6-9$\times$ larger agreement degradation under prompt rephrasings than humans, falling below the generally accepted threshold in usability studies. This indicates that single-prompt evaluation is insufficiently robust for model ranking. Finally, we show that multi-prompt aggregation substantially improves robustness by emphasizing judgments that remain consistent across prompt variants.

cs.CL↗

arXiv2026

SWE-Clar: Can Coding Agents Clarify Incomplete Client Requirements and Deliver Greenfield Software?

Greenfield software development in freelance settings requires coding agents to turn incomplete client requests into working software. This makes requirement clarification essential, yet clients may be unable to provide continuous feedback over long development cycles, and receiving their answers does not guarantee correct understanding of complex requirements. These challenges motivate two benchmark design choices: clarification precedes code implementation, and its effectiveness is assessed through explicit tests of requirement understanding. Guided by these choices, we introduce SWE-Clar, a benchmark grounded in real, high-value client commissions to evaluate requirement clarification and code implementation in this setting. It comprises 100 tasks based on real software development jobs on Upwork, each with at least US\$4,000 in total payments. Each task begins with multi-round interaction between a coding agent and a client agent, followed by separate runs for code implementation with acceptance testing and clarification assessment. For the latter, we construct 512 questions paired with reference answers from real client-freelancer interactions, covering both explicitly stated requirements and those inferred from the conversational context. Both runs begin from the same clarification context but remain isolated, preventing assessment questions from guiding code implementation. Experiments with nine frontier models show that even GPT-6 Astra, which has the highest observed implementation score, achieves a task pass rate of only 67%, underscoring the challenge of independent software delivery in this setting. The dataset and evaluation code are available at https://github.com/upwork/SWE-Clar.

cs.SE↗

arXiv2026

Distribution-Level Contrastive Supervision for Generative Recommendation

Recent generative recommenders improve scalability by retrieving items through token generation instead of traditional ranking over large candidate sets. Yet their training signals are still dominated by discrete code prediction, which overlooks the soft assignment information naturally produced by the tokenizer. This mismatch limits semantic transfer from the tokenizer to the recommender and may hurt overall optimization. We tackle this limitation by introducing a distribution-based supervision scheme for generative recommendation, where multi-level codebook probabilities are treated as soft semantic targets. On top of this design, we develop SODA, a plug-and-play alignment framework that adopts a BPR-style contrastive objective to align recommender representations with target-side distributional representations against negative ones. The proposed method enriches training with finer semantic cues while leaving the decoding stage unchanged. Experimental studies on multiple real-world benchmarks demonstrate that SODA consistently strengthens diverse generative recommendation architectures. Code is available at https://github.com/freyasa/SODA

cs.IR↗

arXiv2026

Reinforcement Learning with Decomposed Subtasks

Group Relative Policy Optimization (GRPO) and related policy-gradient methods for training language model agents collapse an entire multi-turn rollout into a single scalar trajectory reward before it enters the policy update. When the task composes distinct skills, especially under sparse and delayed environmental feedback, this collapsing is lossy: the optimizer must implicitly infer which competency drove the outcome and how that should change behavior. We argue the right primitive is not a better scalar but a decomposition: trajectory reward should be split along subtasks before it enters the policy update. We introduce Reinforcement Learning with Decomposed Subtasks (RLDS), whose core is Subtask-Decomposed Advantage Estimation (SDAE): a replacement for the scalar GRPO advantage that splits trajectory reward into per-subtask shares on a fixed taxonomy, computes a group-relative advantage per subtask, and distributes per-token credit by weighting each subtask's advantage by its importance, concentrating it around the step where a reflection marks that subtask's execution as consequential. We evaluate on four agentic benchmarks: FrozenLake (sparse grid navigation), HotpotQA (multi-hop QA, one retrieval tool), ScienceWorld (long-horizon embodied science), and DeepResearch (long-form research, four tools, composite rubric reward). Heterogeneity diagnostics emitted during training show where decomposition pays off - gains scale with subtask heterogeneity, largest on the high-heterogeneity tasks ScienceWorld (+11.5 points, paired-bootstrap 95% CI [+9.8, +13.3]) and FrozenLake (+9.8 points, [+7.0, +12.8]), and within noise on HotpotQA and DeepResearch, where the diagnostics predicted little to recover. ScienceWorld is also more compute-efficient under RLDS than scalar GRPO (-10.9% wall-clock per step), as long rollouts amortize the fixed reflect-and-grade overhead.

cs.AI↗

arXiv2026

Evaluating Language Models in Realistic Conversational Contexts

As Large Language Models (LLMs) are increasingly deployed to serve open-ended, multi-turn interactions, evaluating conversational quality at human scale has become a central challenge. Existing evaluation frameworks built for summarization, translation, or short-form QA tasks fall short of adequately measuring the consistency of human-scale dialogue, especially when derivation and validation of these metrics themselves often rely on synthetic rather than human sources. We fill the gap by introducing UPHELD (UPwork Human-Scale Evaluated Long Dialogues), a large, reference-full benchmark for evaluating human-scale conversational ability beyond factual correctness. UPHELD consists of hundreds of complete human-to-human dialogues authored by professional script writers, with realistic turn densities and 36,000+ per-turn human annotations across 30,000+ expert-generated dialogue turns. Using UPHELD, we systematically evaluate classical automatic metrics and reference-free LLM-as-a-judge approaches, and find them unreliable when correlated with expert human judgment. Building off this analysis, we use UPHELD to develop a Mixture-of-Judges framework that combines multiple evaluative signals and improves correlation with human assessments by approximately 30%. Overall, UPHELD provides a robust, human-grounded foundation for evaluating human-scale conversational intelligence that fills a crucial gap in the pre-existing LLM dataset landscape.

cs.CL↗

arXiv2026

Bi-Level Optimization for Generative Recommendation: Bridging Tokenization and Generation

Generative recommendation is emerging as a transformative paradigm by directly generating recommended items, rather than relying on matching. Building such a system typically involves two key components: (1) optimizing the tokenizer to derive suitable item identifiers, and (2) training the recommender based on those identifiers. Existing approaches often treat these components separately--either sequentially or in alternation--overlooking their interdependence. This separation can lead to misalignment: the tokenizer is trained without direct guidance from the recommendation objective, potentially yielding suboptimal identifiers that degrade recommendation performance. To address this, we propose BLOGER, a Bi-Level Optimization for GEnerative Recommendation framework, which explicitly models the interdependence between the tokenizer and the recommender in a unified optimization process. The lower level trains the recommender using tokenized sequences, while the upper level optimizes the tokenizer based on both the tokenization loss and recommendation loss. We adopt a meta-learning approach to solve this bi-level optimization efficiently, and introduce gradient surgery to mitigate gradient conflicts in the upper-level updates, thereby ensuring that item identifiers are both informative and recommendation-aligned. Extensive experiments on multiple real-world datasets demonstrate that BLOGER consistently outperforms state-of-the-art generative recommendation methods while maintaining practical efficiency with no significant additional computational overhead, effectively bridging the gap between item tokenization and autoregressive generation. We release our code at https://github.com/Ten-Mao/BLOGER.

cs.IR↗

arXiv2026

UniGRec: Unified Generative Recommendation with Soft Identifiers for End-to-End Optimization

Generative recommendation has recently emerged as a transformative paradigm that directly generates target items, surpassing traditional cascaded approaches. It typically involves two components: a tokenizer that learns item identifiers and a recommender trained on them. Existing methods often decouple tokenization from recommendation or rely on asynchronous alternating optimization, limiting full end-to-end alignment. To address this, we unify the tokenizer and recommender under the ultimate recommendation objective via differentiable soft item identifiers, enabling joint end-to-end training. However, this introduces three challenges: training-inference discrepancy due to soft-to-hard mismatch, item identifier collapse from codeword usage imbalance, and collaborative signal deficiency due to an overemphasis on fine-grained token-level semantics. To tackle these challenges, we propose UniGRec, a unified generative recommendation framework that addresses them from three perspectives. UniGRec employs Annealed Inference Alignment during tokenization to smoothly bridge soft training and hard inference, a Codeword Uniformity Regularization to prevent identifier collapse and encourage codebook diversity, and a Dual Collaborative Distillation mechanism that distills collaborative priors from a lightweight teacher model to jointly guide both the tokenizer and the recommender. Extensive experiments on real-world datasets demonstrate that UniGRec consistently outperforms state-of-the-art baseline methods. Our codes are available at https://github.com/Jialei-03/UniGRec.

cs.IR↗

arXiv2025

UpBench: A Dynamically Evolving Real-World Labor-Market Agentic Benchmark Framework Built for Human-Centric AI

As large language model (LLM) agents increasingly undertake digital work, reliable frameworks are needed to evaluate their real-world competence, adaptability, and capacity for human collaboration. Existing benchmarks remain largely static, synthetic, or domain-limited, providing limited insight into how agents perform in dynamic, economically meaningful environments. We introduce UpBench, a dynamically evolving benchmark grounded in real jobs drawn from the global Upwork labor marketplace. Each task corresponds to a verified client transaction, anchoring evaluation in genuine work activity and financial outcomes. UpBench employs a rubric-based evaluation framework, in which expert freelancers decompose each job into detailed, verifiable acceptance criteria and assess AI submissions with per-criterion feedback. This structure enables fine-grained analysis of model strengths, weaknesses, and instruction-following fidelity beyond binary pass/fail metrics. Human expertise is integrated throughout the data pipeline (from job curation and rubric construction to evaluation) ensuring fidelity to real professional standards and supporting research on human-AI collaboration. By regularly refreshing tasks to reflect the evolving nature of online work, UpBench provides a scalable, human-centered foundation for evaluating agentic systems in authentic labor-market contexts, offering a path toward a collaborative framework, where AI amplifies human capability through partnership rather than replacement.

cs.AI↗

arXiv2025

GraphMatch: Fusing Language and Graph Representations in a Dynamic Two-Sided Work Marketplace

Recommending matches in a text-rich, dynamic two-sided marketplace presents unique challenges due to evolving content and interaction graphs. We introduce GraphMatch, a new large-scale recommendation framework that fuses pre-trained language models with graph neural networks to overcome these challenges. Unlike prior approaches centered on standalone models, GraphMatch is a comprehensive recipe built on powerful text encoders and GNNs working in tandem. It employs adversarial negative sampling alongside point-in-time subgraph training to learn representations that capture both the fine-grained semantics of evolving text and the time-sensitive structure of the graph. We evaluated extensively on interaction data from Upwork, a leading labor marketplace, at large scale, and discuss our approach towards low-latency inference suitable for real-time use. In our experiments, GraphMatch outperforms language-only and graph-only baselines on matching tasks while being efficient at runtime. These results demonstrate that unifying language and graph representations yields a highly effective solution to text-rich, dynamic two-sided recommendations, bridging the gap between powerful pretrained LMs and large-scale graphs in practice.

cs.LG↗

arXiv2025

Delayed Feedback Modeling with Influence Functions

In online advertising under the cost-per-conversion (CPA) model, accurate conversion rate (CVR) prediction is crucial. A major challenge is delayed feedback, where conversions may occur long after user interactions, leading to incomplete recent data and biased model training. Existing solutions partially mitigate this issue but often rely on auxiliary models, making them computationally inefficient and less adaptive to user interest shifts. We propose IF-DFM, an \underline{I}nfluence \underline{F}unction-empowered for \underline{D}elayed \underline{F}eedback \underline{M}odeling which estimates the impact of newly arrived and delayed conversions on model parameters, enabling efficient updates without full retraining. By reformulating the inverse Hessian-vector product as an optimization problem, IF-DFM achieves a favorable trade-off between scalability and effectiveness. Experiments on benchmark datasets show that IF-DFM outperforms prior methods in both accuracy and adaptability.

cs.LG↗

arXiv2020

Atlas: End-to-End 3D Scene Reconstruction from Posed Images

We present an end-to-end 3D reconstruction method for a scene by directly regressing a truncated signed distance function (TSDF) from a set of posed RGB images. Traditional approaches to 3D reconstruction rely on an intermediate representation of depth maps prior to estimating a full 3D model of a scene. We hypothesize that a direct regression to 3D is more effective. A 2D CNN extracts features from each image independently which are then back-projected and accumulated into a voxel volume using the camera intrinsics and extrinsics. After accumulation, a 3D CNN refines the accumulated features and predicts the TSDF values. Additionally, semantic segmentation of the 3D model is obtained without significant computation. This approach is evaluated on the Scannet dataset where we significantly outperform state-of-the-art baselines (deep multiview stereo followed by traditional TSDF fusion) both quantitatively and qualitatively. We compare our 3D semantic segmentation to prior methods that use a depth sensor since no previous work attempts the problem with only RGB input.

cs.CV↗

arXiv2020

DELTAS: Depth Estimation by Learning Triangulation And densification of Sparse points

Multi-view stereo (MVS) is the golden mean between the accuracy of active depth sensing and the practicality of monocular depth estimation. Cost volume based approaches employing 3D convolutional neural networks (CNNs) have considerably improved the accuracy of MVS systems. However, this accuracy comes at a high computational cost which impedes practical adoption. Distinct from cost volume approaches, we propose an efficient depth estimation approach by first (a) detecting and evaluating descriptors for interest points, then (b) learning to match and triangulate a small set of interest points, and finally (c) densifying this sparse set of 3D points using CNNs. An end-to-end network efficiently performs all three steps within a deep learning framework and trained with intermediate 2D image and 3D geometric supervision, along with depth supervision. Crucially, our first step complements pose estimation using interest point detection and descriptor learning. We demonstrate state-of-the-art results on depth estimation with lower compute for different scene lengths. Furthermore, our method generalizes to newer environments and the descriptors output by our network compare favorably to strong baselines. Code is available at https://github.com/magicleap/DELTAS

cs.CV↗

arXiv2020

SuperGlue: Learning Feature Matching with Graph Neural Networks

This paper introduces SuperGlue, a neural network that matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs are predicted by a graph neural network. We introduce a flexible context aggregation mechanism based on attention, enabling SuperGlue to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from image pairs. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on the task of pose estimation in challenging real-world indoor and outdoor environments. The proposed method performs matching in real-time on a modern GPU and can be readily integrated into modern SfM or SLAM systems. The code and trained weights are publicly available at https://github.com/magicleap/SuperGluePretrainedNetwork.

cs.CV↗

arXiv2020

MagicEyes: A Large Scale Eye Gaze Estimation Dataset for Mixed Reality

With the emergence of Virtual and Mixed Reality (XR) devices, eye tracking has received significant attention in the computer vision community. Eye gaze estimation is a crucial component in XR -- enabling energy efficient rendering, multi-focal displays, and effective interaction with content. In head-mounted XR devices, the eyes are imaged off-axis to avoid blocking the field of view. This leads to increased challenges in inferring eye related quantities and simultaneously provides an opportunity to develop accurate and robust learning based approaches. To this end, we present MagicEyes, the first large scale eye dataset collected using real MR devices with comprehensive ground truth labeling. MagicEyes includes $587$ subjects with $80,000$ images of human-labeled ground truth and over $800,000$ images with gaze target labels. We evaluate several state-of-the-art methods on MagicEyes and also propose a new multi-task EyeNet model designed for detecting the cornea, glints and pupil along with eye segmentation in a single forward pass.

cs.CV↗

arXiv2020

Scan2Plan: Efficient Floorplan Generation from 3D Scans of Indoor Scenes

We introduce Scan2Plan, a novel approach for accurate estimation of a floorplan from a 3D scan of the structural elements of indoor environments. The proposed method incorporates a two-stage approach where the initial stage clusters an unordered point cloud representation of the scene into room instances and wall instances using a deep neural network based voting approach. The subsequent stage estimates a closed perimeter, parameterized by a simple polygon, for each individual room by finding the shortest path along the predicted room and wall keypoints. The final floorplan is simply an assembly of all such room perimeters in the global co-ordinate system. The Scan2Plan pipeline produces accurate floorplans for complex layouts, is highly parallelizable and extremely efficient compared to existing methods. The voting module is trained only on synthetic data and evaluated on publicly available Structured3D and BKE datasets to demonstrate excellent qualitative and quantitative results outperforming state-of-the-art techniques.

cs.CV↗

arXiv2019

Efficient 2.5D Hand Pose Estimation via Auxiliary Multi-Task Training for Embedded Devices

2D Key-point estimation is an important precursor to 3D pose estimation problems for human body and hands. In this work, we discuss the data, architecture, and training procedure necessary to deploy extremely efficient 2.5D hand pose estimation on embedded devices with highly constrained memory and compute envelope, such as AR/VR wearables. Our 2.5D hand pose estimation consists of 2D key-point estimation of joint positions on an egocentric image, captured by a depth sensor, and lifted to 2.5D using the corresponding depth values. Our contributions are two fold: (a) We discuss data labeling and augmentation strategies, the modules in the network architecture that collectively lead to $3\%$ the flop count and $2\%$ the number of parameters when compared to the state of the art MobileNetV2 architecture. (b) We propose an auxiliary multi-task training strategy needed to compensate for the small capacity of the network while achieving comparable performance to MobileNetV2. Our 32-bit trained model has a memory footprint of less than 300 Kilobytes, operates at more than 50 Hz with less than 35 MFLOPs.

cs.CV↗