Mohammad Reza Rezaei, Dingxian Wang, Brett Levert, Celeste Fejzo, Andrew Rabinovich, Zhao Chen
Before an AI agent takes on a paid job, someone has to judge whether it is likely to succeed. We test whether frontier language models can make this call before any attempt. We propose PAYWORTHY which contains 1{,}494 real freelance job postings, sampled from 18.8M postings made between 2020 and 2026 on a freelancing platform. An agent attempts each job from the posting alone in a single run, and an LLM judge acting as the client decides whether to accept the work. We ask 18 models to forecast how likely the agent is to succeed and measure how well each model's forecasts rank the accepted jobs above the rejected ones (AUC, where 0.5 is a random ranking). The answer depends on how the agent is described. Told what the agent cannot do, every model ranks the jobs well (AUC 0.66--0.86). Told instead that it is a generic agent with a browser and code execution, the same models rank the same jobs about as well as a random ordering (0.45--0.56). These results hold under judges from four other model families (Claude, GLM, Kimi and Qwen), and the informed forecasts also rank the verdicts of two human raters. Code is available at https://github.com/upwork/Payworthy
Sajjad Pakdamansavoji, Mohammad Reza Rezaei, Ilija Subasic, Pedro Garzon, Radu Paradovschi, Zhao Chen
Language models in customer-facing products must be continually updated, since the data they serve shifts over time. As retraining on the full accumulated data is impractical, each update should instead run on a small set of examples carrying the training signal of all data seen so far. The strongest current methods build this set by gradient matching, which synthesizes examples whose training gradients match those of the full data. However, they require ground-truth labels, run once on a fixed snapshot rather than on a growing stream, and are slow on large data. We introduce \textbf{OLGA}, which lifts these limitations. It distills unlabeled data by drawing the gradient matching targets from the model's own next-token predictions. Across five text classification benchmarks with phi-1.5, it improves accuracy by 10 percentage points over the strongest selection baseline using as few as five examples. OLGA also distills online data streams by keeping a running rehearsal memory of synthetic examples, which is re-distilled together with each new batch as it arrives. In the online setting, it outperforms the strongest selection-based continual baseline by 6.6 stream-AUC percentage points. Finally, our method scales to large data by pairing distillation with selection in a hybrid pipeline, letting selection handle the mild stage of compression and distillation the extreme one. It cuts runtime by 30\% versus distillation alone at similar accuracy, and beats the best selection-only pipeline by up to 12.6 percentage points.
Mohammad Reza Rezaei, Dingxian Wang, Nikhil Verma, Sajjad Pakdamansavoji, Andrew Rabinovich, Zhao Chen
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
Ilija Subasic, Brett Levert, Andrew Rabinovich, Sajjad Pakdamansavoji, Zhao Chen
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.
Large language models (LLMs) perform well on structured reasoning tasks, but what they encode and whether encoded information guides their behavior remain unclear. We introduce a controlled auditing framework that uses parametric CAD to distinguish failures of encoding, expression, and activation-level control. Applied to eight pretrained models and matched randomly initialized controls, our framework reveals that pretraining contributes substantially more to linear decoding of local relations than to global constraint status, which is already highly decodable from random representations. Yet strong local decodability does not translate into the models' zero-shot predictions, which remain far below supervised probe performance. Restoring clean internal activations can repair probe predictions, whereas simple activation steering does not provide reliable targeted control. With only 400 labeled relation pairs, low-rank fine-tuning substantially improves generation on three representative models, showing that the representation--behavior gap can be reduced through adaptation. Together, these findings show why decodability alone is insufficient evidence that encoded information will be expressed or controllable in model behavior.