Publications

Large Language Models (LLMs) trained on extensive scientific research are increasingly integrated as assistants for scientific discovery. However, most research papers omit the fine-grained cognitive process of examining constraints, failed alternatives, and iterative decisions required to achieve the desired goal. Such cognitive processes are vital for real-world scientists working toward specific goals under constraints. In this paper, we show that LLMs, when trained to produce such cognitive traces, perform better as scientific discovery assistants than when trained solely on scientific literature. We propose COGTRL, a trajectory-level reinforcement learning framework that trains LLMs to emulate cognitively grounded reasoning by jointly optimizing cognitive traces and the scientific steps produced in an interleaved manner. Across two 3B-parameter models and two scientific domains (AI and Materials Science), COGTRL improves method quality by an average of 7.85 points over comparable 3B model baselines and achieves competitive performance relative to 70B parameter models. Moreover, analysis by domain experts shows a preference for methods generated by COGTRL over the baselines.

Conventional vulnerability analysis relies on source-level or binary-level access, or dynamic interaction, all of which may be unavailable to third-party analysts auditing closed-source, remotely hosted, or commercially gated software. Therefore, we propose a new paradigm of no-box vulnerability analysis in which neither source code nor runtime interaction is accessible, and only functionality metadata is available. Such metadata defines the intended behavior of the system, including its inputs, outputs, and side effects, while constraining the space of implementations consistent with that behavior. The intended behavior implies irreducible data flows, the minimal source-to-sink data flow skeletons shared by all conforming implementations. By reasoning over irreducible data flows, an analyst can formulate vulnerability hypotheses without observing or interacting with the target system. The analyst can later validate these hypotheses when additional access becomes available. We showcase the feasibility of no-box vulnerability analysis through implementing a pipeline called MCPSEC, which audits Model Context Protocol (MCP) servers for indirect prompt injection vulnerabilities using only the tool metadata exposed at server registration time. We evaluate MCPSEC on 20 widely deployed MCP servers comprising 177 tools, among which human evaluators confirm 95 vulnerable tools. MCPSEC identified 143 tools as vulnerable, and for each vulnerable tool, it produces Theory of Concepts (ToCs), a hypothesized attack scenario for later analyst validation. Using metadata alone, MCPSEC recovers 94 (98.9% recall) real verified vulnerabilities, compared with 80 (84.2% recall) for the LLM baseline. Overall, our results introduce no-box vulnerability analysis as a new analysis paradigm and demonstrate its practical feasibility in realistic systems.

Shows that mid-training language models on their own self-generated data improves downstream reinforcement learning, yielding stronger reasoning than RL alone.

Decouples exploration from generation at inference time so repeated sampling produces genuinely diverse candidate solutions, improving pass@50 by ~21.6% over standard repeated sampling.

A challenging, realistic benchmark of diverse real-world open-source software for evaluating LLM agents on automatically compiling projects, plus OSS-Build-Agent, a strong baseline with an enhanced build-instruction retrieval module.

Shows that as LLMs get better at reasoning they become more susceptible to novel jailbreaks. Introduces ACE and LACE — attacks that encode malicious queries with custom and layered ciphers — and CipherBench to measure cipher-decoding ability.

A GRPO-based interactive training method where a teacher model gives corrective feedback on a student model’s rollouts, instilling self-reflective reasoning without distillation.

Combines multi-agent simulation with an evolutionary algorithm for query rewriting: multiple LLM agents act as simulated shoppers, and their averaged scores form a dynamic reward that iteratively refines the query — improving fitness by ~22%.

A diagnostic benchmark spanning eight domains that evaluates LLMs on six dimensions of reasoning about actions and change, including new ramification constraints for indirect effects. State-of-the-art models struggle across all dimensions, especially ramifications.

A contamination-free, time-sensitive QA benchmark built on synthetic facts, forcing models to do genuine temporal reasoning rather than recalling pre-training knowledge.

Goal-driven, constraint-guided LLM agents that generate scientific hypotheses for materials discovery and design.

Systematically constructs contexts that break common assumptions and shows that, while models reason well over assumption-following contexts, performance drops by up to 20% when those assumptions are broken.