California: Google AI and Peking University researchers unveiled PaperBanana Thursday, a multi-agent framework automating academic diagram generation for research papers.
The system addresses a critical bottleneck in scientific publishing. While AI scientists handle literature reviews and code, visualizing complex discoveries remains labor-intensive.
PaperBanana orchestrates five specialized agents across two phases. The Linear Planning Phase deploys Retriever, Planner, and Stylist agents. The Iterative Refinement Phase uses Visualizer and Critic agents across three improvement rounds.
The Retriever Agent identifies 10 relevant reference examples from databases. The Planner Agent translates technical methodology text into detailed figure descriptions. The Stylist Agent ensures outputs match conference aesthetics like the “NeurIPS Look”.
The Visualizer Agent generates visuals using Nano-Banana-Pro for diagrams. For statistical plots, it writes executable Python Matplotlib code. The Critic Agent inspects images against source text, identifying factual errors or visual glitches.
Researchers introduced PaperBananaBench, a dataset of 292 test cases from NeurIPS 2025 publications. PaperBanana outperformed baselines by 17% overall score, 37.2% conciseness improvement, 12.9% readability gains, and 6.6% aesthetics enhancement.
The system excels in Agent and Reasoning diagrams, achieving 69.9% overall scores. For statistical plots, code-based generation ensures 100% data fidelity versus image models prone to numerical hallucinations.
Domain-specific aesthetic preferences vary significantly. Agent and Reasoning papers favor illustrative 2D vector robots and chat bubbles. Computer Vision research uses camera cones and point clouds. Generative Learning employs 3D cuboids for tensors. Theory papers maintain minimalist grayscale palettes.
The framework is available on GitHub with full documentation.
Mumbai: Wellness startup WellWith has secured ₹1.25 crore in a seed funding round led by BeyondSeed, with participation from Winner Ventures.
The company said the newly raised capital will be channelled into expanding clinical research efforts and strengthening its supply chain infrastructure in Ladakh, a key sourcing region for its products.
Founded in 2021 by Udit Chawla, Rohit Bhavsar, and Nikhil Pratap Singh, WellWith builds wellness solutions derived from sea buckthorn, combining traditional knowledge with modern scientific validation. Its offerings span immunity, digestion, skin health, energy, and preventive wellness, with a focus on clean formulations, minimal processing, and ethical sourcing practices.
With this funding, WellWith plans to undertake structured clinical studies to scientifically validate the safety and effectiveness of its formulations. Alongside research, the startup aims to deepen its on-ground capabilities in the Himalayan region by improving sourcing, processing, and logistics, while ensuring long-term engagement with local communities.
The company claims steady revenue growth so far, largely driven by repeat customers and strong organic demand. It has also built a pan-India direct-selling network designed to promote livelihood creation, supported by community-led initiatives such as its Studentpreneur and Mompreneur programs.
WellWith said its long-term vision is to build a scalable, impact-driven wellness brand rooted in responsible growth, transparent practices, and sustainable value creation for both consumers and sourcing communities.
New Delhi: A new AI-driven trend is gaining momentum across social media platforms, with users turning to ChatGPT to create personalised caricatures from their photos. Unlike traditional photo filters or generic cartoon effects, these AI-generated caricatures reflect individual professions, lifestyles and personal traits, making them feel both playful and deeply personal.
The trend involves users uploading a photograph and prompting the AI to create a caricature-style image. The output typically features lightly exaggerated facial details while keeping the person easily recognisable.
What sets the trend apart is the level of customisation. Many users ask the AI to add context from their everyday lives, journalists appear with notebooks and coffee mugs, designers with sketchpads, and tech professionals against laptop-filled backgrounds. The result is a visual that feels closer to a personality portrait than a simple cartoon.
Ease of use has played a major role in the trend’s rapid spread. Creating an AI caricature requires no artistic skills or specialised software, just a photo and a short prompt. The images are generated quickly, encouraging users to experiment and share the results online.
These caricatures are now being widely used as profile pictures on platforms such as Instagram, LinkedIn, WhatsApp and X. As more users showcase their AI-generated avatars, the ChatGPT caricature trend continues to grow, highlighting how generative AI is reshaping personal expression on social media.
Below are a few ready-to-use GPT Caricature Prompts for you to begin with:
Prompt:“Create a detailed, high-quality caricature based on the uploaded photo. Keep the face clearly recognisable with slightly exaggerated features in a clean, modern cartoon style.
Place the person sitting at an office desk in a professional workspace. The desk should include a laptop, external monitor, keyboard, mouse, pen stand with pens, a notebook, a coffee mug, and a smartphone. Add subtle desk clutter to make it feel realistic but organised.
The person should be dressed in office-appropriate clothing and appear focused yet relaxed, with a confident, friendly expression.
Use soft lighting, warm tones, and a polished illustration style suitable for a LinkedIn or profile picture. The background should show a modern office environment with shelves, documents, and minimal décor.
The overall look should be playful but professional, not overly exaggerated, and visually clean.”
Prompt:“Create a vibrant, high-quality caricature based on the uploaded photos of three friends. Keep all faces clearly recognisable with slightly exaggerated features in a clean, modern cartoon style.
Place the three friends together at a lively party setting. They should be standing or sitting close, laughing and enjoying the moment. Include party elements such as string lights, balloons, confetti, a music speaker, and a decorated table with drinks and snacks. One friend can be holding a drink, another mid-laugh, and the third striking a playful pose.
Dress them in stylish party outfits with different colours and textures to show individual personalities. Their expressions should feel joyful, energetic, and natural—not stiff or posed.
Use warm lighting with soft glows and vibrant colours to capture a festive evening vibe. The background can be a rooftop party, house party, or lounge-style setting with blurred lights for depth.
Keep the illustration playful, colourful, and polished, suitable for Instagram or WhatsApp sharing, without over-exaggeration.”
Overall, the growing use of AI-generated caricatures highlights how people are turning to generative tools to express their identities in more personal and creative ways online. What started as a fun experiment is quickly becoming a new form of digital self-representation, blending individuality with share-ready visuals. As AI tools become more accessible and customisable, such trends are likely to further reshape how users present themselves across social and professional platforms.
Bengaluru: IIT Madras Research Park has partnered with Unicorn India Ventures to launch a Rs 600 crore deep-technology fund focused on backing early-stage, IP-driven startups in engineering-heavy domains.
Named IIT Madras Unicorn Frontier Fund I, the fund was unveiled at the Entrepreneurship Summit 2026 held at the Indian Institute of Technology Madras campus. The fund includes a Rs 400 crore greenshoe option and plans to invest in over 25 startups in its first phase.
The new fund will primarily support startups operating in areas such as robotics, space technology, defence technology, and semiconductors. Portfolio decisions will be jointly led by IIT Madras Research Park and Unicorn India Ventures. While a large share of investment opportunities is expected to come from the IIT Madras ecosystem, the fund will also look at promising deep-tech startups across India.
The announcement was made by V Kamakoti, Director of IIT Madras, during the summit’s opening session.
As per the fund’s structure, initial investments are expected to range between Rs 8–10 crore per startup. The fund will focus on companies at Technology Readiness Levels (TRL) 3 to 4, indicating early validation of core technology. About 60% of the corpus will be deployed in first cheques, while the rest will be reserved for follow-on rounds. The fund has a 10-year tenure, with an option to extend by two years.
The initiative aims to help deep-tech startups emerging from research and academic environments scale their innovations.
Natarajan Malupillai , Chief Executive Officer – IIT Madras Research Park
The fund was announced in the presence of Swapnil Jain, co-founder of Ather Energy and an alumnus of IIT Madras.
IIT Madras Research Park said the fund is part of a broader effort to build India’s capabilities in strategic technology areas by providing patient capital to research-led startups.
SAN FRANCISCO: Anthropic releases Claude Opus 4.6 model Thursday. The timing comes days after Cowork plugins triggered historic software selloff. New model introduces PowerPoint integration and enhanced reasoning capabilities. AI penetration into knowledge work expands significantly.
The model targets office productivity workflows directly. Native Microsoft PowerPoint support enters research preview. Users can generate presentations while maintaining corporate design templates. Claude parses existing layouts and typography automatically.
Opus 4.6 determines when complex requests require extended reasoning. Simple queries get immediate responses. This addresses key weakness in prior iterations. Anthropic claims model outperforms OpenAI’s GPT-5.2 on knowledge work benchmarks. Finance and legal domains show strongest improvements.
Coding improvements enable task distribution across agent teams. Multiple agents mirror human engineering collaboration patterns. Sequential single-agent execution becomes obsolete. Head of Product Management Dianne Penn frames release as inflection point.
Release arrives amid mounting skepticism regarding AI investment returns. Tuesday saw 6 percent software ETF decline. This marked worst single-day performance since April. Thomson Reuters shares plummeted 15.83 percent Tuesday. LegalZoom dropped nearly 20 percent same session.
Industry-specific Cowork plugins sparked the initial panic. Legal research, financial analysis, sales automation faced existential threats. Marketing analytics and data processing tools also vulnerable. Enhanced capabilities renew concerns about specialized software displacement.
Software sector experienced seventh consecutive day of losses. Multiple Big Tech earnings reports project combined $500 billion capex. Anthropic’s aggressive product velocity positions company at debate center. Question remains whether AI spending represents productivity revolution or speculative bubble.
New York: Amazon Web Services (AWS) ended 2025 on a strong note, delivering its fastest quarterly growth rate in more than three years, even as investor concerns weighed on parent company Amazon’s stock.
The cloud business reported $35.6 billion in revenue for the fourth quarter of 2025, up 24% year-on-year, marking its strongest growth in 13 quarters. Amazon said AWS is now operating at an annualised revenue run rate of $142 billion. Operating income for the unit also climbed to $12.5 billion, compared with $10.6 billion in the same quarter last year.
“It’s very different having 24% year-over-year growth on $142 billion annualized run rate than to have a higher percentage growth on a meaningfully smaller base, which is the case with our competitors,” said Andy Jassy during Amazon’s fourth-quarter earnings call. “We continue to add more incremental revenue and capacity than others, and extend our leadership position.”
AWS’s quarterly performance was supported by new deals with companies and government bodies including Salesforce, BlackRock, Perplexity, and the U.S. Air Force.
“More of the top 500 U.S. startups use AWS as their primary cloud provider than the next two providers combined,” Jassy said. “We’re adding significant easy to core computing capacity each day.”
The company added more than one gigawatt of power capacity to its global data centre network during the quarter. Jassy noted that demand continues to come from enterprises migrating from on-premise infrastructure, alongside rising AI workloads.
“We consistently see customers wanting to run their AI workloads where the rest of their applications and data are,” Jassy said. “We’re also seeing that as customers run large AI workloads on AWS, they’re adding to their core AWS footprint as well,” he added.
AWS contributed 16.6% of Amazon’s total $213.4 billion revenue in the quarter. Despite the strong cloud performance, Amazon shares fell 10% in after-hours trading after the company missed earnings expectations and outlined plans to significantly increase capital spending.
Bengaluru: Background verification startup OnGrid has acquired Reczee, an AI-led recruitment platform, to combine hiring and verification into a single, end-to-end workflow. The companies announced the deal on Thursday, though financial terms were not disclosed.
With the acquisition, OnGrid plans to move background verification earlier in the hiring journey, instead of conducting checks only after a job offer is made. The integration is expected to help employers identify risks sooner and improve trust across the recruitment process.
“Recruitment is where trust begins. Joining OnGrid lets us connect AI-driven hiring with what comes next; onboarding, verification, and long-term workforce accountability via eLockr,” said Raj Patel, founder of Reczee.
Reczee provides an AI-powered recruitment system that handles candidate sourcing, screening, coordination and hiring on a single platform.
Founded in 2016, OnGrid works with more than 4,000 organisations across services such as background verification, identity checks, KYC and KYB. The company claims to have processed over one billion verifications and transactions so far.
“We started out as customers of Reczee, using the platform for hiring at OnGrid, and quickly came to love the product. We’re now genuinely excited about the AI capabilities that Reczee brings to the table and believe this combination will unlock meaningful efficiencies and significantly reduce hiring risk for our clients,” said Piyush Peshwani, co-founder of OnGrid.
Reczee will continue operating with its existing product roadmap, while aligning more closely with OnGrid’s broader workforce infrastructure ecosystem.
SAN FRANCISCO: GitHub has officially announced the expansion of its Agent HQ platform by adding Anthropic’s Claude and OpenAI Codex. This move reinforces GitHub’s broader effort to make AI agents a native part of everyday software development.
No additional subscription is required as the integration is available to Copilot Pro+ and Copilot Enterprise customers. Developers can work with multiple AI coding agents directly inside GitHub, GitHub Mobile, and Visual Studio Code.
Developers can start agent sessions and assign work to Claude, Codex, or GitHub Copilot from issues, pull requests, the Agents tab in enabled repositories, and the agent sessions view in VS Code.
Agent HQ supports a multi-agent approach where the same task can be assigned to different agents. This will allow teams to compare how Copilot, Claude, and Codex reason through architectural tradeoffs, edge cases, and implementation strategies.
By keeping agent interactions within existing workflows, developers can move from idea to implementation more easily. They can use different agents for different steps without switching tools or losing context.
Anthropic highlighted the benefit of meeting developers inside their existing collaboration spaces. “We’re bringing Claude into GitHub to meet developers where they are,” said Katelyn Lesse, Head of Platform at Anthropic. “With Agent HQ, Claude can commit code and comment on pull requests, enabling teams to iterate and ship faster and with more confidence.”
OpenAI also emphasized alignment with this vision. Alexander Embiricos from OpenAI added: “We share GitHub’s vision of meeting developers wherever they work, and we’re excited to bring Codex to GitHub and VS Code.”
GitHub confirmed that access to Claude and Codex will expand to additional Copilot subscription tiers. It is also working with Google, Cognition, and xAI to expand the Agent HQ ecosystem across GitHub, VS Code, and the Copilot CLI.
SAN FRANCISCO: Nvidia-backed Voice AI startup ElevenLabs secures $500 million Series D funding at $11 billion valuation, more than tripling its worth from $3.3 billion just one year ago as enterprise adoption accelerates across conversational AI platforms.
Sequoia Capital led the investment, with partner Andrew Reed joining the board to steer global scaling efforts. Power players like Andreessen Horowitz boosted their position fourfold, ICONIQ tripled their commitment, and fresh capital flowed from Lightspeed Venture Partners, Evantic Capital, and BOND – building on Nvidia’s earlier support.
Cumulative funding now stands at $781 million since the 2022 launch. ElevenLabs has advanced from speech-to-text, real-time dubbing, music generation, and intelligent conversational AI agents to Eleven v3 Conversational model.
The numbers tell a growth story for the ages: ElevenLabs surpassed $330 million in annual recurring revenue (ARR) by late 2025, vaulting from $100 million in under two years. Big-name clients, including Deutsche Telekom, Revolut, Square, and the Ukrainian Government are powering this momentum and merging accessible creator tools with robust enterprise platforms.
Co-founders Mati Staniszewski and Piotr Dabkowski plan to channel proceeds into “ElevenAgents” and “ElevenCreative,” fusing audio innovation with video and proactive AI capabilities.
ElevenLabs keeps pushing its global footprint, now spanning key hubs like London, New York, San Francisco, Warsaw, Dublin, Tokyo, Seoul, Singapore, Bengaluru, Sydney, São Paulo, Berlin, Paris, and Mexico City.
For investors and industry watchers, this positions ElevenLabs as an IPO frontrunner. Amid 2025’s AI funding wave in the U.S. Projections hint at $700 million ARR by mid-2026. Potentially commanding 28x revenue multiples as voice interfaces redefine human-tech interaction. With hubs sprouting in 14 cities, enterprise adoption should accelerate, cementing speech AI’s role in a trillion-dollar agent economy.
Tech portfolios eyeing the next big shift can’t ignore this voice revolution.
Paris: Mistral AI, high-performance AI company introduced Voxtral Transcribe 2, the company’s second-generation speech-to-text models aimed at delivering faster, more accurate, and lower-cost transcription. The release includes two variants, Voxtral Mini Transcribe V2 for batch processing and Voxtral Realtime for live audio applications with latency configurable down to sub-200 milliseconds.
credits: mistral ai
In a post on X, the company described it as “next-gen speech-to-text” offering state-of-the-art transcription, speaker diarization, and sub-200ms real-time latency.
Introducing Voxtral Transcribe 2, next-gen speech-to-text models by @MistralAI. State-of-the-art transcription, speaker diarization, sub-200ms real-time latency. Details in 🧵 pic.twitter.com/0IeiJOpiAZ
The Paris-based AI firm said the new models are designed to compete directly with leading transcription services while significantly reducing costs. Voxtral Mini Transcribe V2 is priced at $0.003 per minute for batch jobs, which the company says is roughly one-fifth the cost of competing offerings such as ElevenLabs’ Scribe v2.
According to Mistral’s internal benchmarks, the models deliver about a 4%-word error rate on the FLEURS dataset, outperforming several well-known transcription systems while also processing audio up to three times faster than some rivals. The company added that the real-time model can match batch-level accuracy at higher latency settings suitable for live subtitling, while lower latency modes introduce only a small increase in error rates.
Voxtral Mini Transcribe V2 includes features such as speaker diarization, word-level timestamps, and context biasing that allow users to add up to 100 domain-specific terms for improved accuracy. Voxtral Realtime, meanwhile, is built for voice agents, live captioning, and call-center automation.
Notably, Voxtral Realtime is released under the Apache 2.0 license, allowing organizations to deploy it on-premises without relying on external APIs. With a 4-billion-parameter footprint capable of running on edge devices, the models are positioned for industries with strict data-privacy requirements, including healthcare and finance.
