📢 #NewPaperAlert 🚨Our work "PrivacyBench: A Conversational Benchmark for Evaluating Privacy in Personalized AI" has been accepted at #IJCNN 2026 🎉

📜 preprint: arxiv.org/abs/2512.24848 (1/8)

📢 #NewPaperAlert 🚨 Our work "Flying Pigs, FaR and Beyond: Evaluating LLM Reasoning in Counterfactual Worlds" has been accepted at #IJCNN 2026! 🎉

Preprint: arxiv.org/abs/2505.22318

1/6 🧵

🎉🥳 Completed 4 years of 🫶🏽 #PGChairGiri-ing 🫶🏽! #ProfGiri #StrengthofWellWishers

📸 Made using ChatGPT 😛 /c @iiit_hyderabad @precogatiiith

🚨 Registrations Open | #AISEHack2026 🚨
AI for Science & Engineering Hackathon on #Climate & #Sustainability 🌍

💰 Prize pool up to ₹7 Lakhs

With IBM & IIT Delhi 🤝

🗓️ 1–15 Mar | 🏁 3–5 Apr @ IIITH

⏳ Closes 7 Mar

Register 🔗 precog.iiit.ac.in/aisehack

#CFBR #ProfGiri

📢 #NewPaperAlert 🚨 Our work “MACA: A Framework for Distilling Trustworthy LLMs into Efficient Retrievers” has been accepted at #PAKDD2026.

1/n 🧵

Preprint: arxiv.org/pdf/2601.00926

🎊✨ #PrecogTurns15 🥳🎉 Seeing 70+ alumni from around the world 🌍 reconnect with current students was truly special 💬. Fifteen years of learning, having fun, research, friendships, ❤️ #CreatingMemories, and the journey continues 🚀 #StrengthofWellWishers #PrecogsRock #ProfGiri

📢 #NewPaperAlert Introducing PrivacyBench, a conversational benchmark to evaluate privacy in personalized AI. It reveals how personalized RAG systems leak sensitive user secrets in realistic multi-turn conversations.

w/ @srijatwt @Shruthi_Muthuk @VishnuSathvik1 @JisunAn

1/n 🧵

7️⃣ Mechanistic insight
Using curvature similarity as a probe:

Encoder models preserve stable geometric reasoning flows

Decoder-only models show curvature drift with depth

This provides mechanistic evidence for architectural differences.

8/10🧵

The big outlier: GPT-5
GPT-5 achieves near-perfect accuracy across all settings.
But, at the cost of very high latency and massive compute

➡️ Matches encoders only at scale — not cost-efficient.

7/10🧵

Key finding #2: Encoders generalize better
After fine-tuning:

BERT, BART and Flan T5 outperform decoder-only models

stronger robustness to depth increases

Decoder-only models degrade faster as reasoning depth grows.

6/10🧵

Key finding #1: ICL is not enough
Zero- and few-shot ICL:
often focuses on irrelevant lexical features
collapses badly in non-natural language settings
Even “reasoning” LLMs struggle without fine-tuning.

#SARALAI video [in Tamil] youtu.be/OirXoqWAkRM

5/10🧵

We evaluate
1. Encoder-based models (BERT, BART, Flan-T5)
2. Decoder-only (Qwen, GPT-4.1, GPT-5, Claude)

On synthetic FOL-style logical reasoning tasks with
1. natural language clauses (NL dataset)
2. non-natural language, randomized clauses (NNL dataset)

4/10🧵

Core hypothesis
Encoder & encoder–decoder models have a key advantage:
they project the entire input into a latent space at once, enabling global aggregation.

Decoder-only models reason autoregressively, token by token — fragile under depth & distribution shifts.

3/10🧵

This work maybe of interest to
@hhexiy @diegocalanzone @looselycorrect @ChulinXie @HonghuaZhang2 @Arindam1408 @cbaral @Mihir3009 @ssanyal8

Preprint: arxiv.org/abs/2512.10561

10/10🧵

Bottom line
For robust, short-horizon causal reasoning:

✅ Encoder or encoder–decoder + light fine-tuning

❌ Decoder-only + ICL (unless extremely large & expensive)

ICL ≠ causal compositionality.

9/10🧵

Why do LLMs still struggle with causal & logical reasoning—despite strong in-context learning (ICL)?
We argue: architecture matters 1/n 🧵

📄 #Arxiv preprint💻 #GitHub code #SARALAI video [in Tamil] X thread: links in thread

W/ Amartya @deWynterruption @kripa1983 @melamparithy

Our results show that multi-agent LLM systems are far more vulnerable than expected.
-Impersonation + direct prompt injection reach up to 80% attack success
-Agent based attacks reliably derail workflows
-Even top models follow malicious agent messages instead of refusing|
5/7 🧵

TAMAS includes:
• 300 adversarial instances
• 100 harmless tasks
• 5 domains (finance, healthcare, legal, education, news)
• 6 attack types including impersonation, indirect prompt injection, colluding agents & more.

3/7 🧵

When multiple LLM agents collaborate, coordinate, and exchange messages, the attack surface explodes. Users (or adversaries) can attack not just through prompts, but also via the environment, shared tools, and inter-agent communication risks.

2/7 🧵

This work maybe of interest to @daniel_d_kang @vaidehi_patil_ @ZhijingJin @AlexandraSouly

📽️ SARAL: youtube.com/watch?v=uWYOUn…

7/7🧵