Kaustubh Dhole

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Yiyou Sun, Xinyang Han, Weichen Zhang et al.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 subfields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP-relevant impact.

Alexis Newton, Kaustubh Dhole

A major shift from skilled to unskilled workers was one of the many changes caused by the Industrial Revolution, when the switch to machines contributed to decline in the social and economic status of artisans, whose skills were dismembered into discrete actions by factory-line workers. We consider what may be an analogous computing technology: the recent introduction of AI-generated art software. AI art generators such as Dall-E and Midjourney can create fully rendered images based solely on a user's prompt, just at the click of a button. Some artists fear if the cheaper price and conveyor-belt speed that comes with AI-produced images is seen as an improvement to the current system, it may permanently change the way society values/views art and artists. In this article, we consider the implications that AI art generation introduces through a post-industrial revolution historical lens. We then reflect on the analogous issues that appear to arise as a result of the AI art revolution, and we conclude that the problems raised mirror those of industrialization, giving a vital glimpse into what may lie ahead.

Rahul Seetharaman, Aman Bansal, Hamed Zamani et al.

Scaling laws are well studied for language models and first-stage retrieval, but not for reranking. We present the first systematic study of scaling laws for cross-encoder rerankers across pointwise, pairwise, and listwise objectives. Across model size and training exposure, ranking quality follows predictable power laws, enabling larger rerankers to be forecast from smaller runs. Using models up to 150M parameters, we forecast 400M and 1B rerankers on MSMARCO-dev and TREC DL. Beyond forecasting, we derive compute-allocation rules from the fitted joint scaling law and compare them with equal-compute checkpoints, showing that retrieval metrics often favor data-heavy scaling, though the recommendation depends on the training objective. The forecasts are accurate and typically conservative, making them useful for planning expensive large-model training. These results provide practical scaling principles for industrial reranking systems, and we will release code and evaluation protocols.

Jiarui Liu, Kaustubh Dhole, Yingheng Wang et al.

Reinforcement learning with verifiable rewards (RLVR) has recently emerged as a promising framework for aligning language models with complex reasoning objectives. However, most existing methods optimize only for final task outcomes, leaving models vulnerable to collapse when negative rewards dominate early training. This challenge is especially pronounced in honesty alignment, where models must not only solve answerable queries but also identify when conclusions cannot be drawn from the given premises. Deductive reasoning provides an ideal testbed because it isolates reasoning capability from reliance on external factual knowledge. To investigate honesty alignment, we curate two multi-step deductive reasoning datasets from graph structures, one for linear algebra and one for logical inference, and introduce unanswerable cases by randomly perturbing an edge in half of the instances. We find that GRPO, with or without supervised fine tuning initialization, struggles on these tasks. Through extensive experiments across three models, we evaluate stabilization strategies and show that curriculum learning provides some benefit but requires carefully designed in distribution datasets with controllable difficulty. To address these limitations, we propose Anchor, a reinforcement learning method that injects ground truth trajectories into rollouts, preventing early training collapse. Our results demonstrate that this method stabilizes learning and significantly improves the overall reasoning performance, underscoring the importance of training dynamics for enabling reliable deductive reasoning in aligned language models.

Kaustubh Dhole

Recent advances in mechanistic interpretability suggest that intermediate attention layers encode token-level hypotheses that are iteratively refined toward the final output. In this work, we exploit this property to generate adversarial examples directly from attention-layer token distributions. Unlike prompt-based or gradient-based attacks, our approach leverages model-internal token predictions, producing perturbations that are both plausible and internally consistent with the model's own generation process. We evaluate whether tokens extracted from intermediate layers can serve as effective adversarial perturbations for downstream evaluation tasks. We conduct experiments on argument quality assessment using the ArgQuality dataset, with LLaMA-3.1-Instruct-8B serving as both the generator and evaluator. Our results show that attention-based adversarial examples lead to measurable drops in evaluation performance while remaining semantically similar to the original inputs. However, we also observe that substitutions drawn from certain layers and token positions can introduce grammatical degradation, limiting their practical effectiveness. Overall, our findings highlight both the promise and current limitations of using intermediate-layer representations as a principled source of adversarial examples for stress-testing LLM-based evaluation pipelines.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Kaustubh Dhole, Eugene Agichtein · amazon-science

Query Reformulation(QR) is a set of techniques used to transform a user's original search query to a text that better aligns with the user's intent and improves their search experience. Recently, zero-shot QR has been shown to be a promising approach due to its ability to exploit knowledge inherent in large language models. By taking inspiration from the success of ensemble prompting strategies which have benefited many tasks, we investigate if they can help improve query reformulation. In this context, we propose an ensemble based prompting technique, GenQREnsemble which leverages paraphrases of a zero-shot instruction to generate multiple sets of keywords ultimately improving retrieval performance. We further introduce its post-retrieval variant, GenQREnsembleRF to incorporate pseudo relevant feedback. On evaluations over four IR benchmarks, we find that GenQREnsemble generates better reformulations with relative nDCG@10 improvements up to 18% and MAP improvements upto 24% over the previous zero-shot state-of-art. On the MSMarco Passage Ranking task, GenQREnsembleRF shows relative gains of 5% MRR using pseudo-relevance feedback, and 9% nDCG@10 using relevant feedback documents.

Vishwanath Seshagiri, Siddharth Balyan, Vaastav Anand et al.

Logging is a critical function in modern distributed applications, but the lack of standardization in log query languages and formats creates significant challenges. Developers currently must write ad hoc queries in platform-specific languages, requiring expertise in both the query language and application-specific log details -- an impractical expectation given the variety of platforms and volume of logs and applications. While generating these queries with large language models (LLMs) seems intuitive, we show that current LLMs struggle with log-specific query generation due to the lack of exposure to domain-specific knowledge. We propose a novel natural language (NL) interface to address these inconsistencies and aide log query generation, enabling developers to create queries in a target log query language by providing NL inputs. We further introduce ~\textbf{NL2QL}, a manually annotated, real-world dataset of natural language questions paired with corresponding LogQL queries spread across three log formats, to promote the training and evaluation of NL-to-loq query systems. Using NL2QL, we subsequently fine-tune and evaluate several state of the art LLMs, and demonstrate their improved capability to generate accurate LogQL queries. We perform further ablation studies to demonstrate the effect of additional training data, and the transferability across different log formats. In our experiments, we find up to 75\% improvement of finetuned models to generate LogQL queries compared to non finetuned models.

Chase Fensore, Kaustubh Dhole, Joyce C Ho et al.

We present our submission to the LiveRAG Challenge 2025, which evaluates retrieval-augmented generation (RAG) systems on dynamic test sets using the FineWeb-10BT corpus. Our final hybrid approach combines sparse (BM25) and dense (E5) retrieval methods and then aims to generate relevant and faithful answers with Falcon3-10B-Instruct. Through systematic evaluation on 200 synthetic questions generated with DataMorgana across 64 unique question-user combinations, we demonstrate that neural re-ranking with RankLLaMA improves MAP from 0.523 to 0.797 (52% relative improvement) but introduces prohibitive computational costs (84s vs 1.74s per question). While DSPy-optimized prompting strategies achieved higher semantic similarity (0.771 vs 0.668), their 0% refusal rates raised concerns about over-confidence and generalizability. Our submitted hybrid system without re-ranking achieved 4th place in faithfulness and 11th place in correctness among 25 teams. Analysis across question categories reveals that vocabulary alignment between questions and documents was the strongest predictor of performance on our development set, with document-similar phrasing improving cosine similarity from 0.562 to 0.762.

Ashish Shrivastava, Kaustubh Dhole, Abhinav Bhatt et al.

Despite end-to-end neural systems making significant progress in the last decade for task-oriented as well as chit-chat based dialogue systems, most dialogue systems rely on hybrid approaches which use a combination of rule-based, retrieval and generative approaches for generating a set of ranked responses. Such dialogue systems need to rely on a fallback mechanism to respond to out-of-domain or novel user queries which are not answerable within the scope of the dialog system. While, dialog systems today rely on static and unnatural responses like "I don't know the answer to that question" or "I'm not sure about that", we design a neural approach which generates responses which are contextually aware with the user query as well as say no to the user. Such customized responses provide paraphrasing ability and contextualization as well as improve the interaction with the user and reduce dialogue monotonicity. Our simple approach makes use of rules over dependency parses and a text-to-text transformer fine-tuned on synthetic data of question-response pairs generating highly relevant, grammatical as well as diverse questions. We perform automatic and manual evaluations to demonstrate the efficacy of the system.