Naiyarah Hussain

WiCoDE26 – ACM-W Asia-Pacific Celebration of Women in Computing ✨ Held on February 16–17, 2026 at BITS Pilani, Dubai Campus, WiCoDE26 convened global ACM-W leadership, faculty, industry professionals, and 96 registered participants from across the UAE The two-day celebration featured leadership addresses, a structured Career Pathways panel, the WiCoDE26 Ideathon Challenge, cross-institutional collaboration, and focused discussions on ACM-W scholarships, mentorship, and chapter growth. WiCoDE26 strengthened regional alignment within ACM-W Asia Pacific and established a strong foundation for scaling future celebrations in the region. Read the full recap here: https://lnkd.in/d8YCV_6R #WiCoDE26 #ACMW #WomenInComputing #LeadershipInTech #ACMWAsiaPacific BITS Pilani Dubai Campus Birla Institute of Technology and Science, Pilani ACM-W at BITS Pilani, Dubai Elakkiya R ACM-W Europe Birla Institute Of Technology and Science, Pilani Dubai

42

The Bitter Lesson (AI text detection edition): Opus 5.5 and GPT 6 Astra are better and *much cheaper* at detecting AI generated text and human generated text compared to Pangram Labs and GPTZero. https://lnkd.in/gjq-Acbs

15

1 Comment

"Instead of limiting applications, the focus should be on enabling access to quality datasets, compute infrastructure, and GPUs so that startups can build effectively." Our Co-Founder and Chief Product and Engineering Officer Bharath Shankar had a lot to say at the India AI Impact Summit - about sovereign AI, India's language diversity, and what it truly takes to build for a billion people. Read the full interview to know more. The numbers will surprise you.

83

2 Comments

Thank you, Daniel, for adding our paper to the list and for putting together all the #HCI and #AI papers.   Here's the preprint of our paper on Co-Designing Collaborative AI Tools for Freelancers. https://lnkd.in/eK_QB7J9

12

AI benchmarks are evolving not just in scale, but in cultural depth. I’ve been contributing to efforts like Global PIQA, which evaluates physical commonsense reasoning across 100+ languages, including Telugu, to better understand how models reason in diverse linguistic and cultural settings. What excites me most is how these multilingual evaluations connect to real systems retrieval, vector search, and agent pipelines that need to understand people in their native context, not just translate English text. We’re entering a phase where language, culture, and reasoning will define the next breakthroughs in AI quality and fairness. Building culturally aware evaluation and ingestion loops isn’t an academic curiosity anymore it’s a requirement for global-scale systems. Grateful to the Global PIQA team especially Tyler Chang Catherine Arnett for the opportunity to contribute and represent Telugu in this important work. I’d love to connect with others working on multilingual retrieval, vector search, and culturally grounded AI let’s share insights! #AI #MultilingualAI #Telugu #GlobalPIQA #NLP #CommonsenseReasoning #Inclusion #Research #OpenAI #MachineLearning #LanguageTechnology #Telugu #VectorSearch #RAG #AIResearch #CulturalIntelligence #Google #HuggingFace

12

Token Efficiency in LLMs: A Critical Factor for Indian Languages 🇮🇳 Thanks to Sarvam AI for highlighting an important but often overlooked technical insight. A small architectural decision can create significant real-world impact. Consider this: The Hindi word “नमस्ते” is tokenized very differently across models: • Many traditional global LLMs → 6 tokens • Indian-language–optimized LLMs → 2 tokens At a surface level, this looks trivial. From a systems perspective, it is not. Why Token Efficiency Matters Tokenization directly influences: • Inference Cost — Fewer tokens reduce compute cycles and API expenditure • Latency — Lower token counts translate to faster response times • Scalability — Efficient processing enables high-volume deployments • Language Fidelity — Better segmentation improves contextual and semantic accuracy For India, where vernacular-first AI adoption is accelerating rapidly, this is strategically significant. Language-optimized architectures can unlock: • Affordable AI solutions for startups • Stronger regional-language customer support systems • Greater digital inclusion for the next billion users AI is not one-size-fits-all. Localization and language-aware optimization will shape the next wave of model innovation. It will be exciting to see how Indian AI ecosystems evolve and compete on the global stage. #AI #LLM #GenerativeAI #IndianAI #TokenEfficiency #AIForIndia #LanguageModels #TechInnovation

26

What if the real bottleneck in LLM progress isn’t reasoning — but memory discipline? A recent paper from Tencent AI Lab, “The Pensieve Paradigm: Stateful Language Models Mastering Their Own Context” (Liu et al., 2026, arXiv:2602.12108), makes a bold claim: today’s LLMs fail not because they can’t think, but because they can’t manage their own state. Current models are fundamentally append-only. Their interaction state grows monotonically until the context window fills up and performance degrades. The authors reformulate this dynamic so the next state is not just concatenation, but a function that can modify what remains visible 2602.12108v1 . In short, the model can prune its own history. StateLM is trained to: build an index → search → read chunk → write note → delete context → repeat. The key is the learned deleteContext behavior. Instead of accumulating tokens endlessly, the model develops a “sawtooth” context profile — actively removing noise while preserving distilled notes The results are hard to ignore. On synthetic long-context retrieval, standard Qwen3 models degrade rapidly beyond their window limits, while StateLM remains robust up to 2M tokens On long-document QA and chat memory tasks, it shows consistent double-digit improvements. On BrowseComp-Plus, a deep research benchmark, the gap is striking: ~5% accuracy for the vanilla baseline versus ~52% for StateLM-14B-RL Now the honest part. This is not magic. Architecturally, it resembles structured RAG with iterative pruning. The real contribution is not the tools — it’s training the model to manage context as a learned capability. The authors also show that simply exposing an instruct model to the same toolkit via prompting does not achieve comparable performance Context management, it turns out, is not a prompt engineering trick. It is a policy learning problem. Is this a full paradigm shift? Not yet. But it’s a serious step toward models that don’t just consume context — they curate it. Paper: Liu et al., 2026. The Pensieve Paradigm: Stateful Language Models Mastering Their Own Context. arXiv:2602.12108  2602.12108v1 Part 2 of my Agents/LLM Research Series #AgentsUnpacked #LLMDeepDive

2

Our Cognizant AI labs under the leadership of Babak Hodjat has come up with a pioneering approach to fine-tune LLMs. The traditional way of fine-tuning with Reinforcement Learning(RL) is fragile and we are researching and developing a superior alternative called Evolution Strategies(ES)

297

7 Comments