He/Him | PhD Candidate @iitdelhi | LLM Coordination | BSc + MSc @univofdelhi | Prev: GRM Mentee @ibmresearch; SDE @jhatkaadotorg; OSS Fellow @mlhacks
New Delhi, India
Joined June 2017
- Tweets3.3K
- Following465
- Followers365
- Likes6.4K
Sudipto Ghosh retweeted
Congratulations to Sudipto Ghosh (@ScientificGhosh) on being awarded the prestigious @IndiaDST #INSPIRE #PhDFellowship! ๐๐ We are proud of his achievement and look forward to his continued contributions and success in the years ahead! โจ
Sudipto Ghosh retweeted
** ๐๐๐๐ฉ๐ญ๐ข๐ฏ๐ ๐๐ ๐๐ง๐ญ ๐๐จ๐จ๐ซ๐๐ข๐ง๐๐ญ๐ข๐จ๐ง **
Multi-agent systems are powerful, but they can drastically multiply inference costs. Many existing systems rely on fixed or densely activated agent pipelines without adapting computation to each query: Which agents actually need to be consulted? How deep should the reasoning go? And when is communication worth its compute cost?
Presenting ๐๐๐๐๐ -- Gated Routing and Adaptive Depth for Efficient Reasoning
๐ย Preprint: arxiv.org/abs/2607.10836
GRADE optimises multi-agent reasoning by:
๐ง ๐๐๐๐ซ๐ง๐๐ ๐๐จ๐จ๐ซ๐๐ข๐ง๐๐ญ๐ข๐จ๐ง:ย We built a hierarchical system governed by lightweight gates that jointly manage agent selection, routing depth, communication, and pruning dynamically per query.
โ๏ธ ๐๐จ๐๐๐๐ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ : We adapt GRPO for collaborative settings with a novel, critic-free RL recipe that assigns a shared advantage signal to all participating agents and gates during a rollout.
๐ ๐๐จ๐ญ-๐๐ฐ๐๐ฉ๐ฉ๐๐๐ฅ๐ ๐๐ฑ๐ฉ๐๐ซ๐ญ๐ฌ: GRADE features an Expert Registry with per-agent calibration maps. You can swap out expert models at inference time using just 64 anchor queries, without retraining the gates.
๐ At ~17B average active parameters, GRADE outperforms all baselines on GSM8K, GPQA, and MMLUPro -- beating the strongest baseline on MMLUPro by 4.8 points while using ~39% fewer active parameters.
w/ @ScientificGhosh
Do check out many more exciting works on small models and agentic coordination being developed as part of our mega project -- ๐๐๐ซ๐๐ฆ๐๐ง๐ฎ
parmanu.lcs2.in/
@lcs2lab @iitdelhi
#LLMEfficiency #MultiagentRounting #AgenticAI
Sudipto Ghosh retweeted
๐จ New #TMLR Paper Alert ๐จ
Our paper "Disentangling Intrinsic Importance from Emergent Structure in Multi-Expert Orchestration" has been accepted to the #TMLR!
๐ฅ @ScientificGhosh @sujoynath_ai @sunnymanchanda @Tanmoy_Chak
๐ Paper: arxiv.org/abs/2602.04291
@TmlrOrg @YardiScAI
Sudipto Ghosh retweeted
๐ฌ Parmanu (Hindi for Atom) is live.
Parmanu is part of the Computational Social Systems (LCS2) @lcs2lab at IIT Delhi, led by Prof. Tanmoy Chakraborty @Tanmoy_Chak , and is our dedicated home for Efficient Large Language Models (LLMs) and Small Language Models (SLMs).
Weโre at a turning point in AI. The future wonโt be defined by scaling alone - it will be shaped by efficiency, accessibility, and real-world deployability. Parmanu is our effort to push this efficiency-first vision forward. โจ๐ค
๐ Explore the project page: parmanu.lcs2.in/
Why Parmanu matters ๐ฅ
โข ๐ A centralized hub for our research, with papers accepted at ICLR, ICML, NeurIPS, TACL, ACL, and TMLR
โข ๐ ๏ธ Open access to tools, code, and artifacts spanning model compression, KV efficiency, PEFT, inference optimization, knowledge distillation, and model coordination
โข ๐ง A growing ecosystem focused on making strong language models smarter per parameter, not just larger
What this means for the community
For researchers ๐ฉโ๐ฌ๐จโ๐ฌ
A curated, evolving resource tied to top-tier venues
Reproducible artifacts and principled problem formulations
A shared space to advance efficiency-centric LLM research
For practitioners ๐ฉโ๐ป๐จโ๐ป
Practical techniques to deploy LLMs under tight latency and memory budgets
Faster paths from paper โ production
Tools that actually work under real deployment constraints
Whatโs coming in 2026 ๐๐ฎ
โข ๐ Efficient LLM/SLM leaderboards
โข ๐งช Open-sourced efficient LLM artifacts
โข โ๏ธ More tools for compression, distillation, and inference
โข ๐ค Deep integration with Hugging Face and other popular libraries
If youโre excited about efficient, sustainable, and scalable AI, check out Parmanu, share feedback, and collaborate with us. The next wave of LLMs wonโt just be bigger - theyโll be leaner, faster, and more impactful. ๐
#EfficientLLMs #SLMs #ModelCompression #InferenceOptimization #KnowledgeDistillation #AIResearch #NLP #ICLR #ICML #NeurIPS #ACL #TACL #TMLR #IITDelhi #LCS2 #Parmanu
Sudipto Ghosh retweeted
๐ Today I gave a keynote at ACM COMPUTE on โRebooting NLP Teaching in the LLM Eraโ.
Slides: linkedin.com/posts/tanmoy-chโฆ
I proposed a 42-hour NLP curriculum that fuses classical statistical NLP with modern LLMs.
I'd love to hear feedback, perspectives, or experiences from the NLP community โ how are you approaching NLP education today? ๐
@Indiaacm #ACMCOMPUTE #ComputingEducation #AI4Education
Sudipto Ghosh retweeted
๐ Introducing an open-access course on #LLMs: Intro & Recent Advances! ๐
- Covered both traditional and recent research in LLMs
- A collaborative effort between industry and academia
- W๐๐๐ฌ๐ข๐ญ๐: lnkd.in/g_GEywQF
- ๐๐๐๐ญ๐ฎ๐ซ๐ ๐ฏ๐ข๐๐๐จ๐ฌ: lnkd.in/gGtmp8vP
Sudipto Ghosh retweeted
It's a blockbuster performance by our lab @lcs2lab at #EMNLP2024. 8 out of 8 papers have been accepted. Our gang is coming to Miami. ๐
@emnlpmeeting @iitdelhi
Sudipto Ghosh retweeted
If it were iPhones that were leaving the factory with explosives inside, the media would be a hell of a lot faster to cotton on to what a horrific precedent has been set today. Nothing can justify this. It's a crime. A crime. And everyone in the world is less safe for it.
Sudipto Ghosh retweeted
"เคชเคขเคผเฅเคเคพ เคเคเคกเคฟเคฏเคพ, เคคเคญเฅ เคคเฅ เคฌเคขเคผเฅเคเคพ เคเคเคกเคฟเคฏเคพ"
Sudipto Ghosh retweeted
This is not just inane but also irresponsible for an influential person to put out. Typical techbro posturing passed off as intellectual inquisitiveness.
Sudipto Ghosh retweeted
As news reports come in on the withdrawal of the Broadcasting Services (Regulation) Bill, 2024, I would like to thank everyone who reported, worked and educated others on its dangerous provisions. This is a broad coalition of diverse individuals, groups and organisations who have stood up and asked the government for transparency and contested political censorship.
My own assessment is that these reports indicate merely a retreat and an official ministerial statement or admission in parliament is necessary for confirmation. However, small victories do matter and itโs important to pause and cheer for public accountability.
Hope that the MIB practices the Pre-Legislative Consultation Process (PLCP) in letter and spirit for future legal proposals. I hope to continue providing a political economy framing and a detailed legal analysis that advances policy and public deliberation.
Sudipto Ghosh retweeted
We are in Bangkok for #ACL2024NLP! ๐ @Tanmoy_Chak and his PhD students @shivam6891 @anwoy_ @_themessier are presenting three papers. Catch us at the conference! ๐งโ๐ป
๐ arxiv.org/abs/2405.10548
๐ arxiv.org/abs/2405.11215
๐ arxiv.org/abs/2406.03953
#NLP #LLMs #CSS #Research #Bangkok
Sudipto Ghosh retweeted
๐ข Media Mention! @Tanmoy_Chak recently participated as a panellist on the Yeh Hai India programme on @UrduDoordarshan ๐บ Watch the full interview to listen to his thoughts on the impact of #AI like #ChatGPT on #AcademicIntegrity youtube.com/watch?v=SyFVbsIAโฆ #AIinEducation #Plagiarism
Sudipto Ghosh retweeted
โจJMIR Mental Health! โจ
Glad to share the acceptance of our work: A significant collaborative effort with @aiims_newdelhi, @aiimsrishi, and @yourdost. Check out our LinkedIn post!
linkedin.com/posts/lcs2lab_mโฆ
#MentalHealth #LLMs #AIResearch #MentalHealthResearch #AIforHealthcare
Sudipto Ghosh retweeted
#Day7 One of the busiest days #ACMIndia #ResponsibleAI #SummerSchool with many external speakers joining us. ๐ Thanks @amlanjo Arun, Bharathi @meghaarora42 Sudarsan & Tanuja for sharing your expertise, including showing live demo (#ResponsibleAI part) of systems. ๐คโจ #ProfGiri
