Dayi Lin

Quantum Computing (QC) has tremendous potential - Starting from first principles creating a computer that follow quantum mechanics make so much sense. Wonderful research will emerge from the combination of AI and QC in the coming years - I highly recommend applying (Consider making a transition to the field if you are not already from QC background)!

4

One of the most exciting collaboration projects Haiqu's team is involved in. Open Quantum Design (OQD) is the non-profit effort out of the University of Waterloo / IQC to build a fully open, full-stack trapped-ion quantum computer - from hardware and control electronics to the software layer. Why I care about this (as someone building quantum software for efficient on-device execution): openness at the stack level is what makes results reproducible and enables cross-layer co-design (control <> compilation <> error mitigation <> algorithms). Check the article link in the first comment. Greg Dick Xanadu Open Quantum Design Unitary Foundation #QuantumComputing #OpenSource #TrappedIons #QuantumHardware #QuantumSoftware #OQD

283

9 Comments

Happy to share that our research into advancing instruction cache replacement policies from the Toronto Heterogeneous Compiler Lab has been accepted to MICRO 2025 (https://lnkd.in/gZzKbXzw). Summary: We observe that conventional compiler profile-guided optimization is able to improve spatial locality of instruction memory using code "temperature" information (execution frequency of different segments of code), but does little to better temporal locality. We design a cache replacement policy that uses a lightweight software-hardware interface to pass "temperature" information from the compiler down to the hardware cache replacement policies in order to reduce instruction cache misses on the most frequently execute code. The technique is called TRRIP (Temperature-Based Re-reference Interval Prediction) and is aimed to be realistically implementable (i.e., no changes to instruction set, minimal hardware overheads). The research will be presented in MICRO 2025 in Seoul, Korea on October 20th in Session 3B: Microarchitecture 1. The accepted paper should be available shortly after. A pre-print of the manuscript is available here: https://lnkd.in/gXx58UmC Credit and acknowledgements to co-authors & colleagues: Nikhil Sreekumar, Prabhdeep Soni, Ali Sedaghati, Fang Su, Bryan Chan, Maziar Goudarzi, Reza Azimi

76

3 Comments

My breakthrough in combinatorial optimization A certain class of combinatorial optimization problems asks for the best arrangement, resp. order of things. Some of these problems are "easy" to solve, like e.g. assigning workers to jobs, which is a bit surprising because for n workers and n jobs there are n! posssible assignments of which the optimal one must be calculated. In contrast the Traveling Salesman Problem that asks for the shortest tour through n cities requires finding the shortest tour among (n-1)!/2 possible tours; so it is much harder to find the shortest tour among half as many as there are ways to assign workers to jobs. Ruminating on what could be the culprit for that paradoxical finding resulted in an efficient heuristic for improving tours that finds the optimal way to change the order of O(n) of the cities, whereas known techniques like k-opt moves are subject to combinatorial explosion. The new heuristic works for symmetric and asymmetric instances ; I have verified that with a python implementation. The next major problem, especially in relation to the TSP is subtour elimination, i.e. to guarantee that all cities can be visited with a single tour. The standard techniques are the subtour elimination constraint which are numerically preferable, but they require restarting the calculation with additional constraints that rule out subtours that have already been encountered, and it may be necessary to add exponentially many of these constraints. Then there are the fixed number of Miller Tucker Zemlin constraints that rule out subtours at the cost of much worse numerics and the need to use directed edges even in the case of symmetric instances. I have managed to find an ILP formulation of the TSP that needs no subtour elimination constraints; I have also checked that with an implementation. The iteresting thing is that the reformulation gemeralizes to other notoriously hard problems, even to the quadratic assignment problem where it allows for a formulation as a pure ILP instead of the integer bilinear problem formulation that is the current method of choice. To be very clear: the reformulation of the TSP does not solve the P vs NP question; the unsolved problem is to calculate solutions in whole numbers efficiently. My plans are to do more intensive testing and eventually publish what I have found; I would classify the current status as "in review".

18

3 Comments

The new Canadian Quantum Champions Program emphasizes benchmarking, not just dollars spent. In frontier technologies, progress usually stalls at measurement. If you can't agree on how to evaluate systems, you can't scale them efficiently or responsibly. Programs that tie funding to benchmarks are rare, and risky, but probably necessary if quantum computing is going to move beyond roadmaps and promises. Curious whether others see benchmarking as a real bottleneck, not just qubit counts.

2

2 Comments

Presented our work titled "Diversified Flow Matching with Translation Identifiability" at ICML 2025, Vancouver. This work introduces an ODE-based framework for transporting a diverse collection of source distributions to their target counterparts via a unified flow. This enables unpaired domain translation (e.g., image to image translation) with translation identifiability. 👉 https://lnkd.in/gdVfxCDV

192

1 Comment

“ IQC has also spun out more than 20 quantum startups and 40 per cent of its researchers are involved in commercialization.” Researchers at the Institute for Quantum Computing (IQC) at the University of Waterloo are doing more than tackling trailblazing theoretical and applied researc… Source: Waterloo News https://buff.ly/CPCUxKB

7

I am excited to share that our paper, "Characterizing state space model (SSM) and SSM-Transformer hybrid language model performance with long context length," has been accepted at ISPASS 2026! An early version of the paper is available in this link: https://lnkd.in/gVdj2Psd? <What is this paper about?> We present a thorough workload characterization study of SSM and SSM-Transformer hybrid models with long contexts, focusing on consumer-grade and mobile/edge devices. <Key Findings> - The performance of SSMs is dominated by newer SSM-specific ops, unlike Transformers, which are dominated by GEMM or non-GEMM counterparts. - Unlike pure SSMs, the bottleneck of hybrid models varies by model.  - Performance penalty on non-GEMM operations is more severe on edge devices than workstation-class machines.  (Please refer to our paper for more insights!) <Why SSM and SSM-Transformer hybrid models?> - Supporting long context has become an important feature for LLMs for recent high-value applications (e.g., coding) and high-quality results with the retrieval of external information. However, Transformer-based LLMs suffer from their super-linear overhead on the context length for long contexts. State space models (SSMs) and SSM-Transformer have been proposed to overcome the limitation of Transformer in scaling the context length, demonstrating their superior capability to support extremely long context (e.g., up to 1M tokens in Nvidia's Nemotron3-Nano). - Also, SSM and SSM-Transformer hybrid models are getting more attention in the industry (Nvidia's Nemotron family, Google's Titans, and so on). <Why characterization study targeting mobile/edge devices?> Due to SSM and SSM-Transformer hybrid models' capability to efficiently support long context language model, we envision their adoption would be pervasive in future smart AI devices (IoT, mobile, and wearable devices) where AI service will lead to useful and high-value use cases. However, their computational characteristics are not well-explored. <Acknowledgements> I appreciate the contributions from incredible co-authors at UC Irvine, Rachid Karami, Haocheng Xu, Sitao Huang. Special thanks to Saptarshi Mitra, who led many efforts as the first author! * Note: This research work has been conducted at UC Irvine.

148

8 Comments

I was pleasantly intrigued and interested to hear this week that a coalition of researchers from EleutherAI, MIT, CMU and University of Toronto curated an 8TB dataset composed entirely of open license or public domain texts. From this, they trained Comma v0.1, a 7B parameter model that achieved performance on par with some of the more well known foundation models like Llama 2-7B and others. The implication of this is that it demonstrates that high-performing LLMs can be built without infringing on copyright. A subject which is certainly pretty hot and likely to get hotter in the immediate future due to the need of transparency of the origin and IP of the training data. It points toward more trustworthy and legally compliant models that support regulatory oversight and full transparency on the origin of the training data. Even if, for the time being, they might not fully compete on all benchmarks with all the large, well known foundation models. This could have a good systemic impact in areas that require specialisation, governance and compliance like education, healthcare, scientific research, legal. https://lnkd.in/er75F_HJ #LLMs #TrainingData #Transparency #CopyrightLaws #AI

8

1 Comment

TELUS and Rakuten shared quite different approaches to build/buy AI, in discussion with Nathan Bell, Kearney. Jaime Tatis said TELUS were focused on the discipline of staying open to different partners, whereas Zoran Stejic said Rakuten is building its own AI in Japanese language, citing that its own models were more cost effective if less powerful than general LLMs. TM Forum #DTWIgnite #DTW25 #ai #telcoai #telecoms #strategy #innovation

29

UBC researchers proposed a silicon chip that can convert microwave signals to optical and vice versa with up to 95% efficiency and minimal noise, preserving quantum entanglement. This breakthrough addresses a key hurdle in building scalable quantum networks, enabling long-distance communication between quantum computers over fiber optics. The chip leverages engineered magnetic defects in silicon and superconducting components, operates at ultra-low power, and is compatible with existing chip fabrication technologies. Though still theoretical, this innovation could accelerate the realization of quantum networks, supporting future applications like unbreakable security and enhanced computing capabilities. https://lnkd.in/duYGMXTm

2

A team of researchers at the University of Waterloo have made a breakthrough in quantum computing that elegantly by-passes the fundamental "no cloning" problem. The research, "Encrypted Qubits can be Cloned," appears in Physical Review Letters. We don’t get a quantum future by inventing faster algorithms alone. We get there by building infrastructure that can actually survive reality. A recent breakthrough shows how quantum information can be safely distributed across multiple servers without ever being copied—effectively pointing toward a quantum cloud: a quantum Dropbox, a quantum Google Drive, a quantum-native storage layer built to obey the laws of physics. This matters because quantum computing breaks one of the core assumptions of classical systems: you can’t copy quantum data. No copy means no backups. No backups means no infrastructure. Until now. What’s emerging is a new idea of resilience—redundancy without duplication, security without replication. That’s not just a physics insight; it’s a blueprint for how the quantum internet, quantum data centers, and quantum-native platforms will actually be built. For founders and CTOs, this is the shift from “Can quantum computing work?” to “How do we architect the quantum stack? The quantum era won’t arrive all at once. It will arrive quietly—one infrastructure primitive at a time. And this is one of those primitives. 😇 https://lnkd.in/e-XYf9uQ

75

7 Comments

I have recently finished reading TRM paper [Samsung AI Lab (SAIL) Montréal] Tiny Recursive Models on ARC-AGI-1: Inductive Biases, Identity Conditioning, and Test-Time Compute and here's what i found: Large transformer models often push 100M–1B+ parameters just to gain incremental reasoning improvements, yet still plateau around ~40% accuracy on structured tasks like Sudoku, maze planning, and ARC-AGI-style problems. TRM, with only ~7M parameters, reaches ~80% accuracy on similar benchmarks by enforcing recursive, step-wise reasoning — effectively delivering 2× the accuracy with a fraction of the scale. That gap is a HUGE DEAL: it shows that structured recursion can outperform brute-force parameter growth on logic-heavy tasks. Tiny Recursive Models (TRM) changed how I think about reasoning architectures. After reading the paper, here’s what stood out to me: 1. Recursion instead of scale TRM shows that you don’t always need massive parameter count to solve reasoning heavy tasks. Instead, you reuse a small network recursively. The same module solves progressively refined subproblems. That’s compute reuse, not brute force scale. 2. Depth via time, not parameters Instead of stacking 100+ layers once, TRM applies a small model multiple times. It trades width for iterative refinement. Think: controlled reasoning steps rather than one-pass prediction. 3. Why this matters for HRM HRM (Hierarchical Reasoning Model) separates: High-level planning (H-module) Low-level execution (L-module) The improvement TRM introduces is structural discipline: Iterative refinement makes the planning phase more stable. Recursive reuse reduces overfitting from large monolithic planners. Deep supervision across steps improves intermediate reasoning signals. In short: TRM strengthens HRM by making reasoning compositional instead of oversized. This feels like a shift from “bigger transformer” → “structured computation.” Scaling laws are not the only path forward. Structured recursion is another axis.

4

🚀 Smarter, Faster RAG: A Policy-Optimized, Memory-Efficient Framework for Test-Time Reasoning 📄 New arXiv print 🔗 Scaling Test-Time Inference with Policy-Optimized, Dynamic Retrieval-Augmented Generation via KV Caching and Decoding https://lnkd.in/dErdTDf7 🔍 The Problem Retrieval-Augmented Generation (RAG) boosts LLMs with external knowledge but suffers from: ❌ Static retrieval schedules ❌ Distractor-induced hallucinations ❌ Linear memory growth in KV caching ❌ Inefficient test-time compute without reward alignment. 🧠 Our Framework: PORAG + ATLAS + CRITIC We propose a three-component architecture that dynamically optimizes retrieval, rewards factual and high-quality outputs, and reduces compute/memory costs — all without altering the pretrained base model at inference. 1️⃣ PORAG: Policy-Optimized Retrieval-Augmented Generation • Built on Group Relative Policy Optimization (GRPO) • Fine-tunes with QLoRA adapters, guided by two specialized reward heads: – Retrieval Fidelity (factual correctness) – Response Quality (fluency, coherence, utility) • Enables single-shot decoding with KL-regularized policy updates for stability 2️⃣ ATLAS: Adaptive Token-Layer Attention Scoring • Detects when to retrieve via Multi-Layer Attention Gradient (MLAG) • Builds what to retrieve via Layerwise Representation Pooling (LRP) • Leverages hybrid attention-embedding relevance, with entropy-aware scaling to manage retrieval load 3️⃣ CRITIC: Cache Reduction via Importance-based Token Inclusion Criteria • Selectively retains high-impact tokens in KV cache • Hybrid importance signal: attention strength, entropy spread, gradient sensitivity • Reduces memory and latency for long-context generation, with adaptive compression ⚙️ Exhaustive Test-Time Inference Scaling — No Fine-Tuning Required Our framework integrates advanced inference-time scaling techniques that allow LLMs to trade off compute, latency, and answer quality — dynamically and intelligently. These include: ✔️ Entropy-Guided Decoding — invoke deeper reasoning under uncertainty ✔️ Monte Carlo Tree Search (MCTS) — strategic decoding via multi-branch simulation ✔️ Mixture of Agents — task-routed decoding using a pool of specialized policies ✔️ Lookahead Decoding — speculative multi-token generation and verification These methods are plugged into our test-time control module, which uses token-level entropy, compute budget, and reward models to dynamically select the optimal strategy for each input. 🧠 Conclusion Our framework enables scalable, retrieval-aware LLM inference by combining reward-optimized generation, introspective retrieval, and memory-efficient decoding. It is: ✔️ Transformer-agnostic ✔️ Plug-and-play for SLMs ✔️ Ready for enterprise-scale, compute-sensitive deployments 📌 Full paper on arXiv: https://lnkd.in/dErdTDf7 🤝 With Akash Das, SHIVAM GUPTA, and Venkataramana Runkana #LLMInference #TestTimeScaling

17

Two days. Two LLM launches. Same pattern. Early February: MiniMax M2.5 Early benchmarks position it close to Claude Opus 4.6 — at reportedly ~20x lower cost. Comparable reasoning. Fraction of the price. Yesterday: GLM-5 Another model narrowing the gap with ChatGPT, Claude, and Gemini. And here’s the real shift. Having an LLM with performance close to the 800-pound gorillas of the arena running on your own infrastructure is no longer theoretical. It’s reality. Yes, there is hardware cost. Yes, there is power and maintenance. But we’re not talking about hyperscaler-only territory anymore. We’re talking about serious reasoning capabilities available to companies willing to invest in a limited GPU setup. News like this now drops almost weekly. The moat isn’t model access anymore. It’s integration, product thinking, and governance. Especially in regulated environments, this changes the equation. The question is no longer: “Are these models good enough?” It’s: “What are we building with them?” #AI #LLM #CTOInsights #RegulatedSoftware #OpenSourceAI #MedTech Image taken from the z.ai blog https://z.ai/blog/glm-5

10

7 Comments

The third round of joint testing of LLMs by the network of AI Safety Institutes (AISIs) was focused on multilingual testing of LLM agents! These tests are led by the Singapore AISI and we, at National Research Council Canada / Conseil national de recherches Canada, contributed to the test on behalf of the Canadian AISI, in two languages, Telugu and Farsi. The results are published today on the sidelines of ICML2025. Blog post: https://lnkd.in/eU3yU8AG Full Report: https://lnkd.in/e8N-hNzQ #llmEvaluation #AISafety #MultilingualEvaluation #AgenticAI

75

4 Comments

Toronto startup Taalas launched the HC1 chip on Feb 19, hard-wiring Llama 3.1 8B into a single silicon die delivering 17,000+ tokens/sec per user. It reportedly outperforms Cerebras and Nvidia by wide margins (~8x), uses ~10× less power, and runs in a 200W air-cooled server. Founded in 2023 by Ljubisa Bajic, Taalas has raised $219M and is targeting applications from appliances to medical devices. Impressive speed. Another question is how easily the chip can be customized for new models, given the rapid iteration of model architectures. https://lnkd.in/ghVfZXde

17