Karime Maamari

Karime Maamari, Fadhil Abubaker, Daniel Jaroslawicz et al.

Schema linking is a crucial step in Text-to-SQL pipelines. Its goal is to retrieve the relevant tables and columns of a target database for a user's query while disregarding irrelevant ones. However, imperfect schema linking can often exclude required columns needed for accurate query generation. In this work, we revisit schema linking when using the latest generation of large language models (LLMs). We find empirically that newer models are adept at utilizing relevant schema elements during generation even in the presence of large numbers of irrelevant ones. As such, our Text-to-SQL pipeline entirely forgoes schema linking in cases where the schema fits within the model's context window in order to minimize issues due to filtering required schema elements. Furthermore, instead of filtering contextual information, we highlight techniques such as augmentation, selection, and correction, and adopt them to improve the accuracy of our Text-to-SQL pipeline. Our approach ranks first on the BIRD benchmark achieving an accuracy of 71.83%.

Rahul Singhal, Pradyumna Tambwekar, Karime Maamari

Prompt engineering is effective but labor-intensive, motivating automated optimization methods. Existing methods typically require labeled datasets, which are often unavailable, and produce verbose, repetitive prompts. We introduce PrefPO, a minimal prompt optimization approach inspired by reinforcement learning from human feedback (RLHF). Its preference-based approach reduces the need for labeled data and hyperparameter tuning-only a starting prompt and natural language criteria are needed. PrefPO uses an LLM discriminator to express pairwise preferences over model outputs and provide feedback to an LLM optimizer, iteratively improving performance. We evaluate PrefPO on 9 BIG-Bench Hard (BBH) tasks and IFEval-Hard, a newly-curated, challenging subset of IFEval. PrefPO matches or exceeds SOTA methods, including GEPA, MIPRO, and TextGrad, on 6/9 tasks and performs comparably to TextGrad on IFEval-Hard (82.4% vs 84.5%). Unlike other methods, PrefPO can optimize in both labeled and unlabeled settings. Without labels, PrefPO closely matches its labeled performance on 6/9 tasks, proving effective without ground truth. PrefPO also improves prompt hygiene: we find existing methods produce prompts 14.7x their original length or with 34% repetitive content; PrefPO reduces these issues by 3-5x. Furthermore, both LLM and human judges rate PrefPO's prompts higher than TextGrad's. Finally, we identify prompt hacking in prompt optimizers, where methods game evaluation criteria, and find PrefPO is susceptible at half the rate of TextGrad (37% vs 86%), generating fewer brittle, misaligned prompts.

Daniel Jaroslawicz, Brendan Whiting, Parth Shah et al.

Production-grade LLM systems require robust adherence to dozens or even hundreds of instructions simultaneously. However, the instruction-following capabilities of LLMs at high instruction densities have not yet been characterized, as existing benchmarks only evaluate models on tasks with a single or few instructions. We introduce IFScale, a simple benchmark of 500 keyword-inclusion instructions for a business report writing task to measure how instruction-following performance degrades as instruction density increases. We evaluate 20 state-of-the-art models across seven major providers and find that even the best frontier models only achieve 68% accuracy at the max density of 500 instructions. Our analysis reveals model size and reasoning capability to correlate with 3 distinct performance degradation patterns, bias towards earlier instructions, and distinct categories of instruction-following errors. Our insights can help inform design of instruction-dense prompts in real-world applications and highlight important performance-latency tradeoffs. We open-source the benchmark and all results for further analysis at https://distylai.github.io/IFScale.

Yenchia Feng, Chirag Sharma, Karime Maamari

Although large language models (LLMs) have advanced rapidly, robust automation of complex software workflows remains an open problem. In long-horizon settings, agents frequently suffer from cascading errors and environmental stochasticity; a single misstep in a dynamic interface can lead to task failure, resulting in hallucinations or trial-and-error. This paper introduces $\textit{Environment Maps}$: a persistent, agent-agnostic representation that mitigates these failures by consolidating heterogeneous evidence, such as screen recordings and execution traces, into a structured graph. The representation consists of four core components: (1) Contexts (abstracted locations), (2) Actions (parameterized affordances), (3) Workflows (observed trajectories), and (4) Tacit Knowledge (domain definitions and reusable procedures). We evaluate this framework on the WebArena benchmark across five domains. Agents equipped with environment maps achieve a 28.2% success rate, nearly doubling the performance of baselines limited to session-bound context (14.2%) and outperforming agents that have access to the raw trajectory data used to generate the environment maps (23.3%). By providing a structured interface between the model and the environment, Environment Maps establish a persistent foundation for long-horizon planning that is human-interpretable, editable, and incrementally refinable.

Karime Maamari, Connor Landy, Amine Mhedhbi

Recent advancements in Text-to-SQL, driven by large language models, are democratizing data access. Despite these advancements, enterprise deployments remain challenging due to the need to capture business-specific knowledge, handle complex queries, and meet expectations of continuous improvements. To address these issues, we designed and implemented GenEdit: our Text-to-SQL generation system that improves with user feedback. GenEdit builds and maintains a company-specific knowledge set, employs a pipeline of operators decomposing SQL generation, and uses feedback to update its knowledge set to improve future SQL generations. We describe GenEdit's architecture made of two core modules: (i) decomposed SQL generation; and (ii) knowledge set edits based on user feedback. For generation, GenEdit leverages compounding operators to improve knowledge retrieval and to create a plan as chain-of-thought steps that guides generation. GenEdit first retrieves relevant examples in an initial retrieval stage where original SQL queries are decomposed into sub-statements, clauses or sub-queries. It then also retrieves instructions and schema elements. Using the retrieved contextual information, GenEdit then generates step-by-step plan in natural language on how to produce the query. Finally, GenEdit uses the plan to generate SQL, minimizing the need for model reasoning, which enhances complex SQL generation. If necessary, GenEdit regenerates the query based on syntactic and semantic errors. The knowledge set edits are recommended through an interactive copilot, allowing users to iterate on their feedback and to regenerate SQL queries as needed. Each generation uses staged edits which update the generation prompt. Once the feedback is submitted, it gets merged after passing regression testing and obtaining an approval, improving future generations.

Karime Maamari, Amine Mhedhbi

Recent advancements in Text-to-SQL have pushed database management systems towards greater democratization of data access. Today's language models are at the core of these advancements. They enable impressive Text-to-SQL generation as experienced in the development of Distyl AI's Analytics Insight Engine. Its early deployment with enterprise customers has highlighted three core challenges. First, data analysts expect support with authoring SQL queries of very high complexity. Second, requests are ad-hoc and, as such, require low latency. Finally, generation requires an understanding of domain-specific terminology and practices. The design and implementation of our Text-to-SQL generation pipeline, powered by large language models, tackles these challenges. The core tenants of our approach rely on external knowledge that we extract in a pre-processing phase, on retrieving the appropriate external knowledge at query generation time, and on decomposing SQL query generation following a hierarchical CTE-based structure. Finally, an adaptation framework leverages feedback to update the external knowledge, in turn improving query generation over time. We give an overview of our end-to-end approach and highlight the operators generating SQL during inference.