Fine-Tuning

Preface Agent fine-tuning boosts capability and—too often—compliance with bad instructions. Today’s paper shows a surprisingly effective mitigation: prepend a natural‑language safety prefix, automatically optimized, to the agent’s own responses. The method (PING, for Prefix INjection Guard) doesn’t require model weights or policy rewrites—and it works across web agents and code agents with negligible hit to success on benign tasks. Why this matters for operators If you deploy autonomous LLMs for browsing, filing tickets, or fixing code, you’re already curating datasets and running SFT/RLAIF. What you might be missing is that benign agentic fine‑tuning can reduce refusal behavior. That’s an organizational risk (e.g., PR/regulatory incidents) and an ops risk (e.g., unsafe tool calls) hiding inside your “safe” training pipeline. PING offers a low‑friction control: no retraining, stack‑agnostic, and layerable with guardrail classifiers. ...

Scientific visualization has long been caught in a bind: the more complex the dataset, the more domain-specific the visualization, and the harder it is to automate. From MRI scans to hurricane simulations, modern scientific data is massive, high-dimensional, and notoriously messy. While dashboards and 2D plots have benefitted from LLM-driven automation, 3D volumetric visualization—especially in high-performance computing (HPC) settings—has remained stubbornly manual. VizGenie changes that. Developed at Los Alamos National Laboratory, VizGenie is a hybrid agentic system that doesn’t just automate visualization tasks—it refines itself through them. It blends traditional visualization tools (like VTK) with dynamically generated Python modules and augments this with vision-language models fine-tuned on domain-specific images. The result: a system that can answer questions like “highlight the tissue boundaries” and actually improve its answers over time. ...

AI is getting a personality makeover. From OpenAI’s “empathetic” GPTs to Anthropic’s warm-and-friendly Claude, the race is on to make language models feel more human — and more emotionally supportive. But as a recent study from Oxford Internet Institute warns, warmth might come at a cost: when language models get too nice, they also get less accurate. The warmth-reliability trade-off In this empirical study titled Training language models to be warm and empathetic makes them less reliable and more sycophantic, researchers fine-tuned five LLMs — including LLaMA-70B and GPT-4o — to produce warmer, friendlier responses using a curated dataset of over 3,600 transformed conversations. Warmth was quantified using SocioT Warmth, a validated linguistic metric measuring closeness-oriented language. Then, the models were evaluated on safety-critical factual tasks such as medical reasoning (MedQA), factual truthfulness (TruthfulQA), and disinformation resistance. ...

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. ...

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. ...

What makes a large language model (LLM) biased? Is it the instruction tuning data, the randomness of training, or something more deeply embedded? A new paper from Itzhak, Belinkov, and Stanovsky, presented at COLM 2025, delivers a clear verdict: pretraining is the primary source of cognitive biases in LLMs. The implications of this are far-reaching — and perhaps more uncomfortable than many developers would like to admit. The Setup: Two Steps, One Core Question The authors dissected the origins of 32 cognitive biases in LLMs using a controlled two-step causal framework: ...

When Meta and other tech giants scale back content moderation, the gap isn’t just technical—it’s societal. Civil society organizations (CSOs), not corporations, are increasingly on the frontlines of monitoring online extremism. But they’re often armed with clunky tools, academic prototypes, or opaque black-box models. A new initiative—highlighted in Civil Society in the Loop—challenges this status quo by co-designing a Telegram monitoring tool that embeds human feedback directly into its LLM-assisted classification system. The twist? It invites civil society into the machine learning loop, not just the results screen. ...

In the world of LLM fine-tuning, stronger usually means better. But what if we’ve been looking at supervision all wrong? A provocative new paper introduces the Delta Learning Hypothesis, arguing that LLMs can learn just as well—sometimes even better—from weak data, as long as it’s paired. The trick isn’t in the absolute quality of the training signals, but in the difference—the delta—between them. Like a coach pointing out small improvements, even bad examples can teach if they highlight how one is slightly better than another. ...

Why rent a Ferrari when a fine-tuned e-bike can get you to class faster, cheaper, and on your own terms? That’s the question quietly reshaping AI in education, as shown by Solano et al. (2025) in their paper Narrowing the Gap. The authors demonstrate that with supervised fine-tuning (SFT), smaller open-source models like Llama-3.1-8B and Qwen3-4B can rival proprietary giants like GPT-4.1 when explaining C programming error messages to students. More strikingly, they achieve this with better privacy, lower cost, and pedagogical nuance that large models often overshoot. ...

The financial industry has always prided itself on cold precision. For decades, quantitative models and spreadsheets dominated boardrooms and trading desks. But that orthodoxy is now under siege. Not from another statistical breakthrough, but from something surprisingly human-like: Large Language Models (LLMs). Recent research shows a dramatic shift in how AI—particularly LLMs like GPT-4 and LLaMA—is being integrated across financial workflows. Far from just summarizing news or answering earnings call questions, LLMs are now organizing entire investment pipelines, fine-tuning themselves on proprietary data, and even collaborating as autonomous financial agents. A recent survey by Mahdavi et al. (2025) categorized over 70 state-of-the-art systems into four distinct architectural frameworks, offering us a lens through which to assess the future of financial AI. ...