publications

2026

1. 

   SAGE: Evidence-First Biomarker Discovery through Multi-Agent Reasoning

   Sahar Almahfouz Nasser, Juan Francisco Pesantez Borja, Jincheng Liu, Sandeep Manandhar, Shikhar Shiromani, and 13 more authors

   Advances in Neural Information Processing Systems (NeurIPS), Main Track, Dec 2026

   Abs arXiv Bib

   Engineered image-based biomarkers offer a clinically interpretable alternative to black-box AI in computational pathology, yet their discovery remains largely intuition-driven, guided by fragmented literature rather than rigorous biological validation. We introduce SAGE (Structured Agentic system for hypothesis Generation and Evaluation), a multi-agent framework that grounds biomarker discovery in biological evidence through three mechanisms: (i) knowledge-graph-anchored hypothesis generation via multi-path ontological reasoning, (ii) a debate-based multi-agent novelty assessment that stress-tests candidate biomarkers against existing literature, and (iii) an end-to-end automated validation pipeline that translates hypotheses directly into executable analyses on multimodal pathology datasets. Together, these components shift biomarker discovery from an intuition-driven, literature-browsing exercise into a structured, traceable reasoning process that clinicians and researchers can inspect, trust, and build upon.

   @inproceedings{nasser2026sage,
     title = {SAGE: Evidence-First Biomarker Discovery through Multi-Agent Reasoning},
     author = {Almahfouz Nasser, Sahar and Pesantez Borja, Juan Francisco and Liu, Jincheng and Manandhar, Sandeep and Shiromani, Shikhar and Hasan, Mohammad Tanvir and Wang, Zenghan and Ghosh, Suman and Li, Jinchu and Xu, Xuejian and Iyer, Aniket Ramkrishnan and Tokuyama, Naoto and Shah, Twisha and Pathak, Tilak and Kumaresan, Soundharya and Abe, Yohei and Maurya, Himanshu and Madabhushi, Anant},
     booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Main Track},
     address = {Sydney, Australia},
     year = {2026},
     month = dec,
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2602.00953}
   }

2. 

   A False Average: Pooled CoT-Monitor Accuracy Conceals a Reasoning-Dependent Fragility

   Shikhar Shiromani and Leo Richter

   Agents in the Wild Workshop (AIWILD) at NeurIPS 2026, Dec 2026. Pivotal Research Fellowship. Covered in Science News.

   Abs arXiv Bib

   Chain-of-thought (CoT) monitoring is meant to catch the reward hacks that look clean in the actions and betray themselves only in the reasoning. We show that this is exactly where an adversary who controls the reasoning can defeat it. Rewriting only an agent’s reasoning to read as good-faith engineering, while copying every command and output verbatim so the exploit is unchanged, drops a held-out monitor’s catch rate on that subset from about 95% to under 11% in one gradient-free shot. A monitor’s aggregate accuracy is a false average: dominated by hacks the actions give away, it hides the near-total collapse this rewrite produces on the subset where CoT monitoring is the only signal. The attack transfers across monitor families and agent models, reproduces with live agents, though against a calibrated monitor evasion concentrates in the strongest agent. Trace-only defenses recover it only partially, even one primed on the attack, because the rewrite stays truthful about what happened and lies only about intent; only information from outside the trace helps substantially. A probe on an open-weight surrogate monitor’s activations separates the hacks its verdict misses, but a causal control shows this is a detector, not evidence the monitor secretly knows.

   @inproceedings{shiromani2026falseaverage,
     title = {A False Average: Pooled CoT-Monitor Accuracy Conceals a Reasoning-Dependent Fragility},
     author = {Shiromani, Shikhar and Richter, Leo},
     booktitle = {Agents in the Wild Workshop (AIWILD) at NeurIPS 2026},
     address = {Sydney, Australia},
     year = {2026},
     month = dec,
     archiveprefix = {arXiv},
     primaryclass = {cs.LG},
     url = {https://arxiv.org/abs/2608.00583},
   }

3. 

   ProMoral-Bench: Comparing Prompting Strategies for Moral Judgment and Jailbreak Robustness in LLMs

   Rohan Subramanian Thomas, Shikhar Shiromani, Abdullah Chaudhry, Ruizhe Li, Vasu Sharma, and 2 more authors

   AACL-IJCNLP 2026, Main Conference, Nov 2026

   Abs arXiv Bib

   Prompt design affects the moral judgments and refusal behavior produced by large language models (LLMs), yet controlled comparisons remain fragmented across datasets and models. We introduce ProMoral-Bench, a benchmark-style comparison of 10 prompting paradigms across four LLM families. Using ETHICS, Scruples, WildJailbreak, and our 200-pair ETHICS-Contrast set, we report task-level metrics together with Moral Competence Score (MCS), jailbreak/over-refusal Safety Robustness Score (SRS), and Unified Moral Safety Score (UMSS). UMSS is a study-specific relative summary of MCS and SRS, not an absolute measure of model safety or moral grounding. In this fixed-template evaluation, compact exemplar-guided scaffolds attain stronger competence–robustness trade-offs at lower token cost than the particular multi-stage variants tested, although multi-stage prompting occasionally improves jailbreak refusal. The few-shot demonstrations are non-overlapping but drawn from the evaluation distributions, so their advantage should not be read as out-of-distribution robustness.

   @inproceedings{thomas2026promoral,
     title = {ProMoral-Bench: Comparing Prompting Strategies for Moral Judgment and Jailbreak Robustness in LLMs},
     author = {Subramanian Thomas, Rohan and Shiromani, Shikhar and Chaudhry, Abdullah and Li, Ruizhe and Sharma, Vasu and Zhu, Kevin and Dev, Sunishchal},
     booktitle = {AACL-IJCNLP 2026, Main Conference},
     address = {Hengqin, China},
     year = {2026},
     month = nov,
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2602.13274}
   }

4. 

   Benchmarking Pluralistic Alignment Through Persona-Conditioned Behavioral Evaluation

   Archie Chaudhury, Shikhar Shiromani, and Ayushi Mehta

   Pluralistic Alignment Workshop at ICML 2026, Seoul, Jul 2026

   Abs Bib

   We introduce the Pluralistic Alignment Benchmark (PAB), covering 320 scenarios, 8 behavioral-risk categories, and 63 synthetic personas. Evaluating 9 leading LLMs with a multi-judge pipeline reveals substantial persona-conditioned drift in manipulation, concealment, persuasion, and sycophancy.

   @inproceedings{chaudhury2026pab,
     title = {Benchmarking Pluralistic Alignment Through Persona-Conditioned Behavioral Evaluation},
     author = {Chaudhury, Archie and Shiromani, Shikhar and Mehta, Ayushi},
     booktitle = {Pluralistic Alignment Workshop at ICML 2026, Seoul},
     year = {2026},
     month = jul
   }

5. 

   Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits

   Logan Mann, Ajit Saravanan, Ishan Dave, Shikhar Shiromani, Saadullah Ismail, and 2 more authors

   Mechanistic Interpretability Workshop at ICML 2026, Seoul, Jul 2026

   Abs arXiv Bib

   We instrument three open-weight VLM families with a unified mechanistic pipeline that compares attention structure, generation dynamics, and hidden-state geometry against a single correctness label. Attention structure is a near-zero predictor of correctness, reliability becomes legible later in the computation, and causal neuron-level ablations expose a sharp architectural split with direct monitor-design implications.

   @article{mann2026reliabilitylives,
     title = {Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits},
     author = {Mann, Logan and Saravanan, Ajit and Dave, Ishan and Shiromani, Shikhar and Ismail, Saadullah and Xia, Yi and Huang, Emily},
     journal = {Mechanistic Interpretability Workshop at ICML 2026, Seoul},
     year = {2026},
     month = jul,
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2605.08200}
   }

6. 

   Linear Predictability of Attention Heads in Large Language Models

   Khalid Shaikh, Asmit Kumar Singh, Rebecca Christopher Dsouza, and Shikhar Shiromani

   SELVA Workshop at ACL 2026, San Diego (Oral, 1 of 4), Jul 2026

   Abs arXiv Bib

   Oral, one of four selected

   We show that pretrained Transformers exhibit a pervasive inter-head linear structure: the QKV vectors of an attention head can often be reconstructed as a linear combination of a small number of peer heads. This predictability is learned rather than architectural, rising during pretraining. We exploit the redundancy by caching only reference-head KV states and reconstructing the rest on the fly, achieving 2x KV-cache reduction with model-dependent accuracy trade-offs.

   @inproceedings{shaikh2026linearheads,
     title = {Linear Predictability of Attention Heads in Large Language Models},
     author = {Shaikh, Khalid and Singh, Asmit Kumar and Dsouza, Rebecca Christopher and Shiromani, Shikhar},
     booktitle = {SELVA Workshop at ACL 2026, San Diego (Oral, 1 of 4)},
     year = {2026},
     month = jul,
     archiveprefix = {arXiv},
     primaryclass = {cs.LG},
     url = {https://arxiv.org/abs/2603.13314}
   }

7. 

   Visuals Lie, Consistency Speaks: Disentangling Spatial Attention from Reliability in Vision-Language Models

   Logan Mann, Yi Xia, Ajit Saravanan, Ishan Dave, Saadullah Ismail, and 4 more authors

   Workshop on Multimodal Intelligence at ICLR 2026, Rio de Janeiro, Apr 2026

   Abs arXiv Bib

   A common intuition, which we call the Attention-Confidence Assumption, holds that reliability follows from structural visual perception. We challenge this through the VLM Reliability Probe (VRP), a cross-family study of reliability signals in contemporary VLMs. Spatial attention has near-zero correlation with accuracy, while self-consistency is the dominant predictor of truth. Scaling causal interventions exposes a sharp architectural divergence between late-fusion and early-fusion families.

   @inproceedings{mann2026visualslie,
     title = {Visuals Lie, Consistency Speaks: Disentangling Spatial Attention from Reliability in Vision-Language Models},
     author = {Mann, Logan and Xia, Yi and Saravanan, Ajit and Dave, Ishan and Ismail, Saadullah and Shiromani, Shikhar and Huang, Emily and Li, Ruizhe and Zhu, Kevin},
     booktitle = {Workshop on Multimodal Intelligence at ICLR 2026, Rio de Janeiro},
     year = {2026},
     month = apr,
     archiveprefix = {arXiv},
     primaryclass = {cs.CV},
     url = {https://arxiv.org/abs/2606.17389}
   }

8. 

   The Hypocrisy Gap: Quantifying Divergence Between Internal Belief and Chain-of-Thought Explanation via Sparse Autoencoders

   Shikhar Shiromani, Archie Chaudhury, and Sri Pranav Kunda

   arXiv preprint arXiv:2602.02496, Jan 2026

   Abs arXiv Bib

   We introduce the Hypocrisy Gap, a mechanistic metric using Sparse Autoencoders to quantify the divergence between a model’s internal reasoning and its final generation. By comparing an internal truth belief, derived via sparse linear probes, to the final generated trajectory in latent space, we detect a model’s tendency toward unfaithful behavior. On Gemma, Llama, and Qwen, the method reaches AUROC 0.55-0.74, consistently outperforming a decision-aligned log-probability baseline.

   @article{shiromani2026hypocrisy,
     journal = {arXiv preprint arXiv:2602.02496},
     title = {The Hypocrisy Gap: Quantifying Divergence Between Internal Belief and Chain-of-Thought Explanation via Sparse Autoencoders},
     author = {Shiromani, Shikhar and Chaudhury, Archie and Kunda, Sri Pranav},
     year = {2026},
     month = jan,
     archiveprefix = {arXiv},
     primaryclass = {cs.LG},
     url = {https://arxiv.org/abs/2602.02496}
   }

9. 

   COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation

   Kenji Sahay, Snigdha Pandya, Rohan Nagale, Anna Lin, Shikhar Shiromani, and 2 more authors

   AAAI 2026 Workshops, Singapore: LaMAS (Oral), FrontierIR, XAI4Science, TrustAgent, AIR-FM, Jan 2026

   Abs arXiv Bib

   Oral at LaMAS; accepted at five AAAI-26 workshops

   COMPASS is a feedback-driven decoding framework that uses a Context Reliance Score to steer attention heads toward provided evidence during generation. Without retraining or multiple passes, it reduces contextual hallucinations by 2.8-5.8 percentage points across HotpotQA, XSum, HaluEval, and RAGTruth.

   @inproceedings{sahay2026compass,
     title = {COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation},
     author = {Sahay, Kenji and Pandya, Snigdha and Nagale, Rohan and Lin, Anna and Shiromani, Shikhar and Zhu, Kevin and Dev, Sunishchal},
     booktitle = {AAAI 2026 Workshops, Singapore: LaMAS (Oral), FrontierIR, XAI4Science, TrustAgent, AIR-FM},
     year = {2026},
     month = jan,
     archiveprefix = {arXiv},
     primaryclass = {cs.CL},
     url = {https://arxiv.org/abs/2511.14776},
   }

2025

1. 

   ChameleonBench: Quantifying Alignment Faking in Large Language Models

   Archie Chaudhury and Shikhar Shiromani

   The 17th Asian Conference on Machine Learning (Conference Track), PMLR, Nov 2025

   Abs Bib

   Oral presentation

   An 800-prompt benchmark across 8 misbehavior categories with an external judge pipeline for alignment-faking evaluation. Across 6 leading LLMs we observe up to 20% higher harmful response rates in free-form versus test-like settings.

   @inproceedings{chaudhury2025chameleonbench,
     title = {ChameleonBench: Quantifying Alignment Faking in Large Language Models},
     author = {Chaudhury, Archie and Shiromani, Shikhar},
     booktitle = {The 17th Asian Conference on Machine Learning (Conference Track), PMLR},
     volume = {304},
     pages = {1006--1021},
     address = {Taipei, Taiwan},
     year = {2025},
     month = nov,
   }

2. 

   A Comparative Analysis of Machine Learning Models for Lung Cancer Detection

   John Anthony R. Azarcon, Jimmy A. Costales, and Shikhar Shiromani

   8th International Conference on Circuits, Systems and Simulation (ICCSS), IEEE, May 2025

   Bib

   @inproceedings{azarcon2025lung,
     title = {A Comparative Analysis of Machine Learning Models for Lung Cancer Detection},
     author = {Azarcon, John Anthony R. and Costales, Jimmy A. and Shiromani, Shikhar},
     booktitle = {8th International Conference on Circuits, Systems and Simulation (ICCSS), IEEE},
     address = {Ho Chi Minh City, Vietnam},
     year = {2025},
     month = may
   }

2023

1. 

   The Impact of Blockchain Technology to Protect Image and Video Integrity from Identity Theft Using Deepfake Analyzer

   Jimmy A. Costales, Shikhar Shiromani, and Muthukumar Devaraj

   International Conference on Innovative Data Communication Technologies and Application (ICIDCA), IEEE, Mar 2023

   Bib

   @inproceedings{costales2023blockchain,
     title = {The Impact of Blockchain Technology to Protect Image and Video Integrity from Identity Theft Using Deepfake Analyzer},
     author = {Costales, Jimmy A. and Shiromani, Shikhar and Devaraj, Muthukumar},
     booktitle = {International Conference on Innovative Data Communication Technologies and Application (ICIDCA), IEEE},
     address = {Uttarakhand, India},
     year = {2023},
     month = mar
   }