Genloop

#TuesdayPaperThoughts Edition 66: CLaRa This week's TuesdayPaperThoughts examines "CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning" from researchers at Apple and The University of Edinburgh. While RAG has been the go to solution for grounding LLMs since 2023, most systems still suffer from a fundamental architectural flaw that retrieval and generation operate in separate worlds. CLaRa proposes a rethink: what if they are unified? Key Takeaways: 1️⃣ Salient Compressor Pretraining (SCP): The framework introduces QA-driven and paraphrase-based data synthesis to train the compressor. Rather than token-by-token reconstruction, SCP focuses on semantic essentials. Compressed representations outperform text-based baselines by 2.36% on Mistral-7B despite using 16× less context. 2️⃣ Weakly Supervised Joint Training: CLaRa trains the query reasoner and generator end-to-end using only next-token prediction loss, with gradients flowing through a differentiable top-k estimator. No relevance labels needed. This approach aligns retrieval relevance with answer quality 3️⃣ Great Results: On HotpotQA with 4× compression, it achieves 96.21% Recall@5, exceeding the fully supervised BGE-Reranker (85.93%) by +10.28% proving that joint optimization with weak supervision can outperform explicit retrieval training. Can CLaRa become a new standard for RAG architectures? Research Credits: Jie He, He Bai, Sinead Williamson, Jeff Z. Pan, Navdeep Jaitly, Yizhe Zhang Paper Link: In comments

Top 3 papers of the Week [Dec 24 - Dec 28, 2025] suggested by Genloop's LLM Research Hub. 🥇 CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning This research introduces CLaRa, a unified framework for retrieval-augmented generation that integrates embedding-based compression and joint optimization in a shared continuous space. By leveraging SCP for data synthesis and end-to-end training, it achieves state-of-the-art performance in QA benchmarks, aligning retrieval relevance with answer quality. Research Credits: Jie He; Richard He Bai; Sinead Williamson; Jeff Z. Pan; Navdeep Jaitly; Yizhe Zhang 🥈 Latent Collaboration in Multi-Agent Systems The research introduces LatentMAS, a training-free framework enabling latent collaboration among LLM agents, enhancing system-level reasoning and efficiency. It achieves higher accuracy, faster inference, and reduced token usage compared to text-based MAS, demonstrating substantial advancements in generative AI. Research Credits: Jiaru Zou; Xiyuan Yang; Ruizhong Qiu; Gaotang Li; Katherine Tieu; Ke Shen; Hanghang Tong; Yejin Choi; Jingrui He; James Zou; Mengdi Wang; Ling Yang 🥉 What does it mean to understand language? The research explores how language understanding involves exporting information from the brain's language system to other regions for constructing mental models, leveraging cognitive neuroscience to study the neural and cognitive basis of comprehension. Research Credits: Colton Casto; Anna Ivanova; Evelina Fedorenko; Nancy Kanwisher Check out the links below for the LLM Research Hub and the top papers

#TuesdayPaperThoughts Edition 65: Souper-Model This week's #TuesdayPaperThoughts examines "Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance" from Meta. This paper introduces SoCE (Soup Of Category Experts), a framework that achieves state-of-the-art results through intelligent model weight averaging. Key Takeaways: 1️⃣ Category-Aware Expert Selection: SoCE analyzes Pearson correlations across benchmark categories to identify weakly-correlated clusters, then selects expert models for each cluster. Rather than uniform averaging, SoCE uses optimized weighted combinations to aggregate these complementary strengths. 2️⃣ State-of-the-Art Results Without Retraining: SoCE achieves 80.68% accuracy on BFCL for 70B models a 2.7% absolute improvement over the previous best model. For 8B models, the gains are even more pronounced at 5.7% relative improvement, reaching 76.50% accuracy. This demonstrates that strategic model souping can unlock performance improvements without the computational expense of retraining from scratch. 3️⃣ Enhanced Model Consistency: Beyond raw performance, souped models exhibit significantly higher Pearson correlations between category performances compared to unsouped counterparts. In large-scale experiments across 800+ checkpoints, souped candidates consistently gained on 20+ out of 36 categories. This suggests souping doesn't just improve aggregate scores it creates more robust, coherent models across diverse task types. SoCE provides both a scientific framework for model aggregation and a practical recipe that democratizes access to state-of-the-art performance transforming how we think about reusing and combining existing checkpoints rather than always training from scratch. Research Credits: Shalini Maiti, Amar Budhiraja, Bhavul Gauri, Gaurav Chaurasia, Anton Protopopov, Alexis Audran--Reiss, Michael Slater, Despoina Magka, Tatiana Shavrina, PhD, Roberta Raileanu, Yoram Bachrach Paper Link: In comments (edited) 

Top 3 papers of the Week [Nov 17 - Nov 21, 2025] suggested by Genloop's LLM Research Hub. 🥇 Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance The paper introduces Soup Of Category Experts (SoCE), a novel model souping method using non-uniform weighted averaging to enhance LLM performance. By leveraging weakly-correlated benchmark categories, SoCE achieves state-of-the-art results in multilingual tasks, tool calling, and math, improving robustness and efficiency. Research Credits: Shalini Maiti; Amar Budhiraja; Bhavul Gauri; Gaurav Chaurasia; Anton Protopopov; Alexis Audran--Reiss; Michael Slater; Despoina Magka; Tatiana Shavrina, PhD; Roberta Raileanu; Yoram Bachrach 🥈 Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs The paper introduces Nemotron Elastic, a framework for reasoning-oriented LLMs with hybrid Mamba-Attention architectures, enabling nested submodels optimized for diverse deployment budgets. It achieves significant cost reductions and state-of-the-art performance through innovations like elastification techniques and multi-budget optimization. Research Credits: Ali Taghibakhshi; Sharath Turuvekere Sreenivas; Saurav Muralidharan; Ruisi Cai; Marcin Chochowski; Ameya Mahabaleshwarkar; Yoshi Suhara; Oluwatobi Olabiyi; Daniel Korzekwa; Mohammad Patwary; M. Shoeybi; Jan Kautz; Bryan Catanzaro; Ashwath Aithal; Nima Tajbakhsh; Pavlo Molchanov 🥉 MMaDA-Parallel: Multimodal Large Diffusion Language Models for Thinking-Aware Editing and Generation The research introduces ParaBench, a benchmark for evaluating text and image outputs, and proposes MMaDA-Parallel, a multimodal diffusion framework optimized with Parallel Reinforcement Learning to enhance cross-modal alignment and semantic consistency, achieving a 6.9% improvement over state-of-the-art models. Research Credits: Ye Tian; Jiongfan Yang; Anran Wang; Yu Tian; Jiani Zheng; Haochen Wang; Zhiyang Teng; Zhuochen Wang; Yinjie W.; Mengdi Wang; Xiangtai Li Check out the links below for the LLM Research Hub and the top papers

Heading to Dubai this week for #IGFMiddleEast2025: NXT Frontiers to speak about delivering reliable, personal, enterprise-ready LLMs. Looking forward to conversations with leaders of fast-moving organizations that are accelerating AI adoption for data-driven decision making. DM me if you’re around and open to a warm chat about #AI, #LLMs, and #Agentic #DataAnalysis — would love to connect. India Global Forum | Genloop

#TuesdayPaperThoughts Edition 64: TiDAR - Think in Diffusion, Talk in Autoregression This week's #TuesdayPaperThoughts examines "TiDAR: Think in Diffusion, Talk in Autoregression" from researchers at NVIDIA. What if you could generate text with the speed of parallel diffusion models and the quality of autoregressive decoding without compromise? The paper proposes a hybrid architecture that achieves exactly this by combining parallel token drafting with autoregressive sampling in a single forward pass. Key Takeaways: 1️⃣ Dual-Mode Architecture in Single Forward Pass: TiDAR operates both diffusion and autoregressive modes simultaneously using structured attention masks. The model drafts multiple tokens in parallel from the marginal distribution while sampling final outputs from the chain-factorized joint distribution via rejection sampling. This exploits "free token slots" on modern GPUs where additional tokens incur minimal latency increase in memory-bound operations essentially getting parallel computation at almost no extra cost. 2️⃣ Quality-Speed Balance: The paper reports 4.71× speedup for TiDAR 1.5B and 5.91× for TiDAR 8B compared to standard autoregressive models while maintaining competitive quality. TiDAR generates 7.45 and 8.25 tokens per forward pass respectively, outperforming both speculative decoding methods like EAGLE-3 and diffusion models like Dream and Llada on coding and math benchmarks. For the first time, a diffusion-based architecture surpasses speculative decoding in measured throughput. 3️⃣ Simplified Training Strategy: TiDAR employs full masking for the diffusion section during training rather than random corruption. This provides denser loss signals, simpler loss balancing between AR and diffusion objectives, and enables one-step diffusion drafting during inference. TiDAR's approach offers a practical path forward for speed and quality text generation. Research Credits: Jingyu Liu, Xin Dong, Zhifan Ye, Rishabh Mehta, Yonggan Fu, Vartika Singh, Jan Kautz, Ce Zhang, Pavlo Molchanov Paper Link: In comments

Top 3 papers of the Week [Nov 10 - Nov 14, 2025] suggested by Genloop's LLM Research Hub. 🥇 TiDAR: Think in Diffusion, Talk in Autoregression The paper introduces TiDAR, a hybrid generative architecture combining diffusion-based drafting and autoregressive sampling, achieving high throughput, GPU efficiency, and AR-level quality. TiDAR outperforms existing models in token generation speed and quality, closing the gap with AR models while maintaining parallelizability. Research Credits: Jingyu Liu; Rishabh Mehta; Yonggan Fu; Vartika Singh; Jan Kautz; Ce Zhang; Pavlo Molchanov 🥈 SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads? The research introduces SWE-fficiency, a benchmark for evaluating automated performance optimization in software repositories, emphasizing code reasoning and long-horizon software engineering. It highlights agent underperformance in localizing bottlenecks and maintaining correctness, aiming to advance automated performance engineering. Research Credits: Jeffrey Ma; Milad Hashemi; Amir Yazdanbakhsh; Kevin Swersky; Ofir Press; Enhui Li; V. Reddi; Parthasarathy Ranganathan 🥉 The Path Not Taken: RLVR Provably Learns Off the Principals The research introduces a parameter-level analysis of RLVR's training dynamics, revealing distinct optimization regimes compared to SFT. It proposes the Three-Gate Theory to explain sparsity artifacts and advocates for RLVR-native, geometry-aware learning algorithms over repurposed SFT methods. Research Credits: Hanqing Zhu; Zhenyu (Allen) Zhang; Hanxian Huang; DiJia Su; Zechun Liu; Jiawei Zhao; Igor Fedorov; Hamed Pirsiavash; Zhizhou Sha; Jinwon Lee; David Z. Pan; Kai Sheng Tai Check out the links below for the LLM Research Hub and the top papers

Cohort 14 of the NetApp Excellerator program just hit a major milestone at Demo Day in Bengaluru, where five game-changing startups — Synthefy, Filo Systems, TrueFoundry, Sentra, and Genloop.ai — unveiled breakthrough innovations in AI, data, and cloud. Over months of collaboration with NetApp’s global experts, these startups sharpened their solutions, unlocked proof-of-concept opportunities, and built scalable go-to-market strategies. Above all, they've pushed the boundaries of intelligent data infrastructure. India’s deep-tech founders are rewriting the rules of enterprise innovation, and we’re proud to stand alongside them. Together, we’re creating an ecosystem where bold ideas thrive: responsibly, inclusively, and at scale. https://bit.ly/3HNwwnw