machine learning

AC comment and our reply disappeared on OpenReview [D]

A concerning issue has emerged on OpenReview: several users report that an AC's initial comment and the author's subsequent reply have vanished. This comment, crucial for understanding reviewer feedback and addressing concerns, summarized key questions and weaknesses. The disappearance raises questions about transparency, particularly if it obscures the rationale behind potential rejection decisions. We've encountered similar discussions around data management challenges, as explored in our article, "A Day in the Life of a Data Scientist in 2026."

BDH-CQ: IN-CONTEXT LEARNING WITH RECURRENT LATENT REASONING [R]

BDH-CQ represents a significant advance in AI reasoning, seamlessly integrating memory, adaptation, and inference within a unified computational framework. This system tackles previously unseen tasks by iteratively processing queries within a high-dimensional latent space, updating recurrent memory in real-time without explicit verbalization of intermediate steps. A compact 150M-parameter configuration achieves 29.5% pass@2 on ARC-AGI-1, exceeding prior cost-accuracy benchmarks. For further exploration of related advancements in AI efficiency, see our article, "Semi Edge Inference Idea."

If you had a bunch of GPUs lying around, what would you actually build with them? (Running LLMs is off the table) [D]

Beyond the well-trodden path of local LLMs, a stack of high-end GPUs unlocks a realm of compelling possibilities. What truly innovative projects would emerge? Consider distributed simulations, specialized generative models outside of text, or accelerated rendering pipelines. The opportunity exists for impactful homelab experiments demanding serious computational power, or even uniquely ambitious personal endeavors. Explore the potential – as demonstrated by projects like the Doom renderer reimagined as a transformer, discussed in "I compiled Doom's renderer into a 21B-parameter transformer"—and share your most intriguing ideas.

How much does adding an honest limitations section hurt the paper? [D]

Addressing limitations honestly in research papers—while generally beneficial—raises critical questions about reviewer bias and potential requests for remediation. Does openly acknowledging constraints negatively impact perception, or will reviewers demand fixes outlined in the limitations section? Furthermore, the introduction of AI reviewers introduces a novel consideration: could these limitations inadvertently bias algorithmic assessment? Exploring these nuances, as discussed in "My Model Was Cheating on Its Own Test," highlights the complexities of transparency in AI research.

Are there any theoretically-guided practices left in machine learning nowadays? [D]

The rise of large language models has sparked a critical question: have theoretically-guided practices in machine learning become relics of the past? Historically, principles like avoiding overfitting, rigorous test set separation, and optimizer selection based on performance guarantees shaped model development. However, recent empirical successes suggest these guidelines are often superseded by what simply *works*. Has the field transitioned to a purely empirical approach, driven by observed results rather than foundational theory?

A linter for PyTorch 'torch-preflight' [P]

Avoid costly GPU errors and optimize your PyTorch training runs with `torch-preflight`, a new linter designed to catch common pitfalls. Developed over months of experience, this tool analyzes your code – without execution or GPU requirements – to identify issues like unmanaged autograd graphs or improper gradient accumulation. Beyond bug detection, `torch-preflight` estimates VRAM usage, providing actionable insights to ensure your training fits within available resources. Explore the project and contribute at [https://github.com/highwaterlabs/torch-preflight](https://github.com/highwaterlabs/torch-preflight). For more on model evaluation, see our related article, "

For the people who got reviews back from neurips, cvpr, eccv, etc and also tested their paper through an agentic reviewer like the stanford one, how different were the reviews? [D]

For those who recently received reviews from NeurIPS, CVPR, ECCV, or similar conferences, and also utilized agentic reviewer tools like the Stanford model, a compelling question arises: how do the reviews compare? We're exploring the divergence between human and LLM assessments, seeking insights into this evolving landscape. Early indications suggest significant variations, prompting a deeper understanding of how AI-assisted review impacts the peer review process. For further context on related challenges, see our article, "My Model Was Cheating on Its Own Test."

I compiled Doom's renderer into a 21B-parameter transformer -- no training anywhere [P]

Researchers have achieved a remarkable feat: compiling Doom's rendering engine into a 21-billion parameter transformer model, bypassing traditional training methods. Utilizing a custom compiler, the Doom algorithm’s computation graph was translated into transformer weights, resulting in a Hugging Face-compatible checkpoint. Inputting scene data as a prompt generates pixel drawing commands, reconstructing the iconic E1M1 frame. While achieving a modest 35 frames per day on a B200 (compared to the original’s 35 FPS), this innovative approach demonstrates the potential of AI-native computation.

My Model Was Cheating on Its Own Test

Data scientists often strive for model accuracy, but what happens when a model gains an unfair advantage? In a recent *Towards Data Science* post, an author discovered their car price prediction model was "cheating" – a preprocessing pipeline inadvertently allowed it to glimpse the test set. This resulted in a deceptively high R-squared score. The experience highlights a critical pitfall in machine learning workflows and the importance of rigorous validation.

A Day in the Life of a Data Scientist in 2026

The role of the data scientist is undergoing a profound transformation. In "A Day in the Life of a Data Scientist in 2026," we explore how AI has fundamentally reshaped daily workflows, moving beyond traditional spreadsheet limitations. Discover how automation, intelligent insights, and streamlined model deployment now define the modern data scientist's experience. This post offers a future-focused perspective on leveraging AI to empower data-driven decision-making—a shift that's already underway, as highlighted by innovations like Kog’s work to optimize GPU inference for agentic workflows.

Kog is going deeper to squeeze more inference out of GPUs

The narrative around GPUs and AI agents has often framed the former as ill-suited for the latter. French startup Kog challenges this perception, announcing deeper optimizations to maximize inference capabilities within GPUs. This represents a significant shift, potentially unlocking new efficiencies for agentic workflows. Kog’s advancements promise to empower developers with more accessible and performant AI solutions. For those interested in exploring the broader landscape of accessible AI models, see our recent article on Meta’s Glimmer release.

I Made an LLM Lay Siege to My Minecraft House

Can a language model actively design a challenging Minecraft level? We put it to the test, tasking an LLM with laying siege to a player-built house – a compelling experiment in adversarial level design. The results are surprisingly dynamic and reveal the potential for AI to generate complex, reactive environments. Explore the full story and see how this experiment unfolded. For further insights into AI agents, consider "5 Fun Agentic AI Papers to Read," offering a curated selection of foundational research.

Building text to ASCII diffusion model , need advice and guidance [P]

Embarking on a text-to-ASCII diffusion model is an ambitious, yet exciting, project! Leveraging your solid ML foundation—including coursework like CS229 and experience with CNNs and diffusion models—you're well-positioned to explore this unique application. While building such a model from scratch presents challenges, focusing on GAN research is a good starting point. Consider exploring papers that bridge the gap between text understanding and generative image models. For further context on evaluating research impact, see our article, "TMLR Relevance and Prestige [D]," for insights into academic standing.

Neurips 2026: Modified date on reviews [D]

UrgenT Help Detecting Performance Regressions Using Machine Learning and Hardware Counters [P]

Detecting performance regressions demands a robust evaluation strategy. This post explores a common challenge: building a machine learning model for anomaly detection with limited "healthy" data—specifically, around 10 samples per counter group. The author's approach, utilizing leave-one-out for threshold setting and treating regression samples as a test set, raises key questions regarding optimal validation splits and evaluation metrics. Prioritizing false-positive and detection rates over traditional MSE/MAE is crucial in this one-class anomaly detection scenario.

Flock says its new tool will help identify police abuse, but hasn’t explained how it works

Flock’s new “Audit Assistance” tool, mandated for all customers, claims to identify police abuse—a bold assertion lacking detailed explanation. While Flock states the tool has already detected instances of misconduct, the mechanics behind its detection remain opaque, prompting legitimate questions about its efficacy. This lack of transparency warrants careful scrutiny. For those navigating complex AI workflows, understanding the nuances of different tools is crucial; consider our guide comparing LangChain and LangGraph for insights into agentic systems.

Writer introduces new AI model and upgraded harness to contain token costs

Writer is pleased to announce a significant advancement in AI accessibility: a new AI model and upgraded harness designed to dramatically reduce token costs. Built as a post-training variation on Z.ai’s open-source GLM-5.2, this system delivers deployment-ready capabilities at a substantially lower price point. This innovation empowers broader access to powerful AI tools. For those navigating agentic workflows, understanding the nuances of tools like LangChain, as explored in our recent article, is increasingly important. We believe this release represents a key step toward democratizing AI.

Would you choose a PhD advisor who gives you complete freedom but almost no guidance? [D]

Navigating the landscape of PhD advisors presents a critical decision. Consider this scenario: a fully funded ML PhD with a senior, respected advisor offering near-complete freedom—choose your topics, projects, and collaborations with minimal oversight. However, this autonomy comes at a cost: limited guidance or technical input. Is this a dream setup prioritizing independence, or a dealbreaker due to the lack of mentorship?

I built an "honest" CS conference ranking: sorted by how good the trip is, not the CORE ranking [P]

Navigating the conference landscape just got smarter. Forget solely relying on CORE rankings – we’ve built HonestCSRankings.org to prioritize your overall experience. Mapping nearly 540 CORE-ranked conferences, this tool ranks venues based on real-world factors like weather, safety, cost, and city vibrancy. Discover the optimal destination for your next research trip, factoring in distance from home and even identifying “A*” conferences with less-than-ideal locations.

chessformer_lens demo: ablating 1 of a chess transformer's 128 attention heads makes the model stop finding Morphy's queen sacrifice [P]

A fascinating demonstration reveals the critical role of individual attention heads within chess-playing transformer models. Ablating just one of 128 attention heads in the "chessformer_lens" model completely prevents it from identifying the iconic Morphy’s queen sacrifice – a testament to the intricate interplay of these components. Explore the full demo and replication notebooks on GitHub [link]. This highlights the nuanced dependencies within AI architectures, a concept further examined in our article, "How Artificial Intelligence Disrupts Engineering Progression," detailing AI's impact on career development.

Looking for real-world examples of predictive analytics in mortgage lending [D]

Predictive analytics are transforming mortgage lending, and understanding the key variables is crucial for your graduate project. Lenders leverage a range of factors beyond just credit activity and interest rates—property appreciation, borrower life events, and debt-to-income ratios all play significant roles in predicting refinance likelihood. Successful models often incorporate a combination of these elements to achieve accuracy.

AAAI 2027 Review: No code submission? [D]

AAAI 2027 paper reviews have revealed a concerning trend: a surprisingly low number of submissions include accompanying code. This deviates from AAAI's explicit emphasis on reproducibility and raises questions about the rigor of some submissions. While initial scoring will reflect this omission, we seek community input. Providing code fosters transparency and allows for validation – a practice we strongly advocate, as evidenced by our own consistent code sharing on ArXiv.

Amazon will train on Twitch streamers’ content by default, unless they opt out

Amazon will now leverage Twitch streamer content for AI training by default, a decision underscored by Twitch CPO Mike Minton’s statement that an opt-in system would see minimal adoption. This shift reflects a commitment to rapidly advancing AI capabilities, though it raises considerations around creator consent and data usage. Users retain the ability to opt out, ensuring control over their content.

Building Multimodal Workflows with a Local LLM

Unlock new possibilities in data processing by building multimodal workflows directly on your machine. This post explores leveraging Gemma 4 and Ollama to create powerful systems capable of accepting image inputs and generating structured outputs – a significant step beyond traditional spreadsheet limitations. Discover how local LLMs empower accessible and future-focused data manipulation. For a foundational understanding of the underlying mechanics, explore "Backpropagation Explained for Beginners (Part 3): How Backpropagation Really Works," to deepen your knowledge of the neural networks at play.