Josefa Lia Stoisser

Josefa Lia Stoisser, Marc Boubnovski Martell, Lawrence Phillips et al.

We propose STRuCT-LLM, a unified framework for training large language models (LLMs) to perform structured reasoning over both relational and graph-structured data. Our approach jointly optimizes Text-to-SQL and Text-to-Cypher tasks using reinforcement learning (RL) combined with Chain-of-Thought (CoT) supervision. To support fine-grained optimization in graph-based parsing, we introduce a topology-aware reward function based on graph edit distance. Unlike prior work that treats relational and graph formalisms in isolation, STRuCT-LLM leverages shared abstractions between SQL and Cypher to induce cross-formalism transfer, enabling SQL training to improve Cypher performance and vice versa - even without shared schemas. Our largest model (QwQ-32B) achieves substantial relative improvements across tasks: on semantic parsing, Spider improves by 13.5\% and Text2Cypher by 73.1\%. The model also demonstrates strong zero-shot generalization, improving performance on downstream tabular QA (TableBench: 8.5\%) and knowledge graph QA (CR-LT-KGQA: 1.7\%) without any QA-specific supervision. These results demonstrate both the effectiveness of executable queries as scaffolds for structured reasoning and the synergistic benefits of jointly training on SQL and Cypher (code available at https://github.com/bouv/STRuCT-LLM).

Josefa Lia Stoisser, Marc Boubnovski Martell, Sidsel Boldsen et al.

As data-science agents shift from co-pilots to auto-pilots, silent misframing becomes a critical failure mode. Agents quietly commit to plausible but unintended task framings, producing clean, executable artifacts that hide their incorrect assessment of the task. Existing benchmarks score whether the pipeline runs, ignoring whether the agent recognized the task was underspecified. We introduce Ambig-DS, two diagnostic suites: one for prediction-target ambiguity (Ambig-DS-Target, 51 tasks built on DSBench, a tabular modeling benchmark) and one for evaluation-objective ambiguity (Ambig-DS-Objective, 61 tasks built on MLE-bench, a Kaggle-style ML competition benchmark), constructed so that scoring uses each source benchmark's original evaluator. For every task we pair the original, fully specified version with an ambiguous variant produced by controlled edits; a human-and-LLM verification pipeline confirms each variant admits multiple plausible interpretations with decision-relevant consequences. The suites are analyzed independently and ambiguity lowers performance in both. Across five agents spanning efficient to frontier-class models, we find in our controlled diagnostic setting: (i) failures are silent commitments: wrong-target submissions on Target, wrong-metric or non-committal baseline submissions on Objective, rather than execution errors; (ii) allowing the agent to ask one clarifying question recovers much of the loss under idealized conditions, suggesting missing framing information drives a substantial part of the observed degradation; but (iii) agents cannot reliably tell when to use it: permissive prompts induce over-asking on clear tasks, while conservative prompts induce silent defaulting on ambiguous ones. Recognizing target and objective underspecification, not pipeline execution, is the bottleneck missing from standard DS-agent evaluations.

Marc Boubnovski Martell, Josefa Lia Stoisser, Kaspar Märtens et al.

Reliable confidence estimation enables safe deployment of chain-of-thought (CoT) reasoning through text-only APIs. Yet the dominant black-box baseline, self-consistency over K samples, is linearly expensive and ignores the geometry of the trace. We propose a black-box trajectory-confidence score: we embed a CoT as a sliding-window trajectory and measure its convergence to external answer anchors with a one-parameter softmax. The method needs no logits, hidden states, or supervised calibrators. Across six (benchmark, reasoner) settings on MedQA-USMLE, GPQA Diamond, and MMLU-Pro with Gemini 3.1 Pro and Claude Sonnet 4.6, fusing this score with coverage and verbalized-confidence channels at K=4 yields Pareto improvements over self-consistency at K=8 in 6/6 settings (median AUC 0.78 vs 0.71, deltaAUC=+0.075). A fixed-pick control (+0.060) and E5 cross-embedder replication rule out answer switching and single-vendor artifacts. Geometry peaks in the penultimate window across benchmarks and reasoners, and inverts at the terminal window on GPQA Diamond. Three unscaffolded regimes separate black-box confidence into a judge-mediated Coverage prior (C), within-trace Geometry (G), and a conditional Verbalization channel (V). Across 18 benchmark x reasoner x proposer settings, C and G provide independent signal in 18/18 and 16/18, while V contributes residual signal in 6/18. Swapping the judge from GPT-5-mini to Claude Sonnet 4.6 leaves G-only AUC unchanged (|delta|<=0.013) and shifts C-only AUC by at most +/-0.02 (kappa=0.82). Fusion beats the best single channel in 17/18 settings (median AUC 0.78, max 0.92).

Josefa Lia Stoisser, Marc Boubnovski Martell, Julien Fauqueur

This work reframes the Text-to-SQL task as a pathway for teaching large language models (LLMs) to reason over and manipulate tabular data--moving beyond the traditional focus on query generation. We propose a two-stage framework that leverages SQL supervision to develop transferable table reasoning capabilities. First, we synthesize detailed chain-of-thought (CoT) traces from real-world SQL queries, providing step-by-step, clause-level supervision that teaches the model how to traverse, filter, and aggregate table fields. Second, we introduce a Group Relative Policy Optimization (GRPO) reinforcement learning objective that connects SQL execution accuracy to generalizable reasoning by encouraging steps that extend beyond task-specific syntax and transfer across datasets. Empirically, our approach improves performance on standard Text-to-SQL benchmarks and achieves substantial gains on reasoning-intensive datasets such as BIRD and CRT-QA, demonstrating enhanced generalization and interpretability. Specifically, the distilled-quantized LLaMA model achieved a relative 33.9\% increase in accuracy when trained on Text-to-SQL tasks, while Qwen achieved a relative 14.5\% increase. These results suggest that SQL can serve not only as a target formalism but also as an effective scaffold for learning robust, transferable reasoning over structured data.

Josefa Lia Stoisser, Marc Boubnovski Martell, Lawrence Phillips et al.

Large language model (LLM) agents are increasingly deployed in structured biomedical data environments, yet they often produce fluent but overconfident outputs when reasoning over complex multi-table data. We introduce an uncertainty-aware agent for query-conditioned multi-table summarization that leverages two complementary signals: (i) retrieval uncertainty--entropy over multiple table-selection rollouts--and (ii) summary uncertainty--combining self-consistency and perplexity. Summary uncertainty is incorporated into reinforcement learning (RL) with Group Relative Policy Optimization (GRPO), while both retrieval and summary uncertainty guide inference-time filtering and support the construction of higher-quality synthetic datasets. On multi-omics benchmarks, our approach improves factuality and calibration, nearly tripling correct and useful claims per summary (3.0\(\rightarrow\)8.4 internal; 3.6\(\rightarrow\)9.9 cancer multi-omics) and substantially improving downstream survival prediction (C-index 0.32\(\rightarrow\)0.63). These results demonstrate that uncertainty can serve as a control signal--enabling agents to abstain, communicate confidence, and become more reliable tools for complex structured-data environments.

Josefa Lia Stoisser, Lawrence Phillips, Aditya Misra et al.

Synthetic chain-of-thought (CoT) traces are widely used to train large reasoning models (LRMs), improving generalization by providing step-level supervision. Yet most approaches require ground-truth labels to seed or filter these traces - an expensive bottleneck in domains like biology where wet-lab data are scarce. We propose a label-free alternative: uncertainty-based filtering, which uses a model's own confidence - quantified through established uncertainty metrics like self-consistency and predictive perplexity - as a substitute for external labels. We sample multiple reasoning traces and retain only low-uncertainty subsets. Applied to biological perturbation prediction, a domain where wet-lab labels are especially costly, we show that the filtered subset has higher accuracy, and that supervised fine-tuning (SFT) on uncertainty-filtered data outperforms unfiltered synthetic data, narrows the gap to ground-truth training, and surpasses strong LRM baselines. Ablations show that per-class filtering corrects for class-specific uncertainty scales and that hybrid uncertainty metrics yield higher-quality datasets. Our results suggest that model-internal confidence is a powerful signal for efficient reasoning dataset creation, enabling LRMs in domains where supervision is expensive.

Lawrence Phillips, Marc Boubnovski Martell, Aditya Misra et al.

Predicting cellular responses to genetic perturbations represents a fundamental challenge in systems biology, critical for advancing therapeutic discovery and virtual cell modeling. While large language models (LLMs) show promise for biological reasoning, their application to perturbation prediction remains underexplored due to challenges in adapting them to structured experimental data. We present SynthPert, a novel method that enhances LLM performance through supervised fine-tuning on synthetic reasoning traces generated by frontier models. Using the PerturbQA benchmark, we demonstrate that our approach not only achieves state-of-the-art performance but surpasses the capabilities of the frontier model that generated the training data. Our results reveal three key insights: (1) Synthetic reasoning traces effectively distill biological knowledge even when partially inaccurate, (2) This approach enables cross-cell-type generalization with 87% accuracy on unseen RPE1 cells, and (3) Performance gains persist despite using only 2% of quality-filtered training data. This work shows the effectiveness of synthetic reasoning distillation for enhancing domain-specific reasoning in LLMs.