Sidhaarth Murali
I’m a student at CMU working on how models learn from feedback, how that feedback can go wrong, and how we can make models better at improving themselves. I also work across the systems stack, from models down to kernels, because I’m interested in the whole loop of how these systems are built and run.
I’m graduating from CMU this December and looking for full-time opportunities to do meaningful research and engineering on misalignment, how people and models exchange feedback, and hermetic long-horizon environments that teach frontier models new capabilities.
Current work
Beyond Parallel Sampling: Diverse Query Initialization for Search Agentspreprint
Parallel search-agent threads issue near-identical first queries, so their failures are correlated. DivInit seeds each thread differently at inference time, with no training.
with João Coelho, Chenyan Xiong
RL with Learnable Textual FeedbackICML 2026
A bilevel RL formulation where the textual feedback given to a model is itself learned, rather than a fixed scalar reward.
with Utsav Singh, Danush Khanna, Souradip Chakraborty, Amrit Singh Bedi
Off-Policy Self-DistillationICLR 2027 submission
When self-distillation stays on-policy the model loses the ability to keep learning. Off-policy rollouts avoid this.
with Andrea Zanette, Aarti Singh, Jeff Schneider
RiskBenchin progress
Whether an agent can recognize risk and hold off an irreversible action until it has enough evidence, instead of guessing.
with Andrea Zanette
RL Environment for Decode Megakernelsposter
An environment where a model learns to fuse an entire decode step into one CUDA kernel on H100, graded hermetically on-device.
@Preference Model
Industry
Preference Model2026
RL environments for frontier-model post-training on GPU kernel authoring. Hermetic evaluation, reward design, roofline baselines.
IBM Research2025
Joint retriever-reader architectures for small language models. Two-stage training with contrastive learning, 60% latency reduction.
Confident AIYC W25
Built DeepTeam, the agentic red-teaming framework in DeepEval. Guardrails for prompt injection, PII leakage, unsafe reasoning.
Systems
I’ve worked with frameworks like VERL and AReaL, trained models with LoRA and FSDP, and used vLLM for inference. I’ve also built environments for training models to optimize CUDA kernels and worked with multi-GPU training setups. I like working across the stack, especially when the systems work helps answer a research question.
Publications
SELF-PERCEPT: Introspection Improves LLM Detection of Multi-Person Mental ManipulationACL 2025
with Aditya Guru, Siddharth Shukla, Tanuj Tyagi, Danush Khanna, Pratinev Seth, Kripa Ghosh
LADDER: Latent Attention and Decomposition for Deep Enhanced Retrieval in Medical QANLDB 2025
with Sowmya Kamath, Supreetha R
ReMAG-KR: Retrieval and Medically Assisted Generation with Knowledge ReductionACL SRW 2024
with Sowmya Kamath, Supreetha R
QA-RAG: Question and Answer-based Retrieved Chunk Re-Formatting for RAGpreprint 2024
with Kaushik Roy, Yuxin Zi, Chathurangi Shyalika, Renjith Prasad, Vedant Palit, Amit Sheth
NBA MVP Prediction and Historical Analysis using Cross-Era ComparisonICCCNT 2024
with Ishan Godbole, Sowmya Kamath
Does Degree Capture It All? Centrality and Clustering in Signed NetworksCODS-COMAD 2024
with Abhin B, Ramya D Shetty, Shrutilipi Bhattacharjee
SCaLAR at SemEval-2024 Task 8: RoBERTa Ensemble for Detecting Machine Generated TextSemEval 2024
with Abhin B, Anand Kumar M
Writing
Nothing here yet.