LLMs

Most tool-augmented LLMs approach math reasoning like they’re wielding a hammer—good for hitting one nail at a time, but ill-equipped when the problem requires a wrench, a compass, and a soldering iron all at once. Enter Multi-TAG, a clever, finetuning-free framework that aggregates the strengths of multiple tools per reasoning step. Think of it as an LLM with a toolbox, not just a single tool. And it doesn’t just work—it wins, posting 6.0% to 7.5% accuracy gains across MATH500, AIME, AMC, and OlympiadBench against top baselines, using both open and closed LLMs. ...

Fine-tuning is the hammer; steering is the scalpel. In an era where models are increasingly opaque and high-stakes, we need tools that guide behavior without overhauling the entire architecture. That’s precisely what GRAINS (Gradient-based Attribution for Inference-Time Steering) delivers: a powerful, interpretable, and modular way to shift the behavior of LLMs and VLMs by leveraging the most fundamental unit of influence—the token. The Problem with Global Steering Traditional inference-time steering approaches often rely on global intervention vectors: a blunt, one-size-fits-all shift in hidden activations derived from paired desirable and undesirable examples. But these methods are insensitive to which specific tokens caused bad behavior. It’s like adjusting a recipe because the dish tastes bad—without checking if the salt or the sugar was at fault. ...

When we talk about parameter-efficient fine-tuning, LoRA (Low-Rank Adaptation) is often celebrated as a silver bullet: cost-effective, memory-efficient, and—many assume—safe. After all, it modifies only a small fraction of model parameters, sideloaded as low-rank matrices, while leaving the massive pretrained model backbone untouched. The prevailing belief has been that such minimal intervention can’t possibly memorize or leak sensitive data. This belief is now decisively debunked by LoRA-Leak, a landmark framework introduced in a new paper by researchers from Tsinghua and HKUST. Their findings are a wake-up call for AI developers and policymakers alike: even LoRA-finetuned models are highly vulnerable to membership inference attacks (MIAs)—and ironically, the very presence of the frozen pretrained model amplifies this leakage risk. ...

Climate science, once defined by hand-tuned code and static diagnostics, is entering a new phase of automation and adaptability. At the forefront is EarthLink, a self-evolving multi-agent AI platform built specifically to support Earth system science. But this isn’t another LLM wrapper for answering climate questions. EarthLink is something deeper: a scientific collaborator that plans experiments, writes code, debugs itself, interprets results, and learns with each use. From Toolkits to Thinking Partners Traditional tools like ESMValTool or ILAMB have standardized climate model evaluation, but they remain brittle and rigid. They require domain-specific programming expertise and offer little flexibility beyond predefined tasks. In contrast, EarthLink introduces a new paradigm: ...

Most large language models (LLMs) are trained as if every piece of data is equally nutritious. But just as elite athletes optimize not just what they eat but when and how they eat it, a new paper proposes that LLMs can perform better if we learn to dynamically adjust their data “diet” during training. The Static Selection Problem Traditional data selection for LLMs is front-loaded and fixed: you decide what data to keep before training, often using reference datasets (e.g., Wikipedia) or reference models (e.g., GPT-3.5) to prune the lowest-quality examples. While effective in reducing cost, this approach ignores a key insight: an LLM’s preference for certain types of data evolves over time. ...

Scientific communication has always suffered from the tyranny of static text. Even the most revolutionary ideas are too often entombed in dense LaTeX or buried in 30-page PDFs, making comprehension an uphill battle. But what if your next paper—or internal training doc—could explain itself through animation? Enter Manimator, a new system that harnesses the power of Large Language Models (LLMs) to transform research papers and STEM concepts into animated videos using the Manim engine. Think of it as a pipeline from paragraph to pedagogical movie, requiring zero coding or animation skills from the user. ...

What if your AI model isn’t just answering questions, but living in its own version of time? A new paper titled The Other Mind makes a bold claim: large language models (LLMs) exhibit temporal cognition that mirrors how humans perceive time — not through raw numbers, but as a subjective, compressed mental landscape. Using a cognitive science task known as similarity judgment, the researchers asked 12 LLMs, from GPT-4o to Qwen2.5-72B, to rate how similar two years (like 1972 and 1992) felt. The results were startling: instead of linear comparisons, larger models automatically centered their judgment around a reference year — typically close to 2025 — and applied a logarithmic perception of time. In other words, just like us, they feel that 2020 and 2030 are more similar than 1520 and 1530. ...

In an age where aging infrastructure meets accelerating AI, a new paper out of George Mason University proposes a novel question: Can large language models interpret what even seasoned engineers find difficult — NDE contour maps of bridges? The answer, based on this pilot study, is a cautious but resounding yes — with caveats that echo through the entire field of AI-assisted engineering. The Problem: Data Is There — Expertise Isn’t Always Bridges are scanned using advanced non-destructive evaluation (NDE) tools — Ground Penetrating Radar (GPR), Electrical Resistivity (ER), Impact Echo (IE), and Ultrasonic Surface Waves (USW) — but interpreting those outputs requires human expertise, which is not always available, especially during emergency assessments or in rural areas. Contour maps from these tools don’t speak for themselves. ...

When it comes to trading the S&P 500, tradition says: trust the chart. But a new study from UCLA researchers proposes a smarter compass—one that listens not only to price momentum but also to the tone of the news. By merging language model-powered sentiment scores with technical indicators and time-series forecasting, the authors build a hybrid strategy that outperforms a buy-and-hold baseline during a volatile 3-month window. ...

What makes code feel like it was written by a student? Not just errors, but how they evolve. Not just style, but how it diverges from the polished norms. This week’s standout paper, ParaStudent, tackles a refreshingly underexplored challenge: teaching LLMs to generate code that learns like a student — messy, iterative, full of hiccups and growth. Instead of building yet another high-performing code assistant, the authors fine-tune LLMs to mimic real students in an introductory CS class at UC Berkeley. They call their framework ParaStudent. The goal: replace idealized solutions with something plausibly human — an LLM that stumbles, recovers, and grows in fidelity to how novices actually write code. ...