MAROKO133 Update ai: Researchers find adding this one simple sentence to prompts makes AI

📌 MAROKO133 Update ai: Researchers find adding this one simple sentence to prompts

One of the coolest things about generative AI models — both large language models (LLMs) and diffusion-based image generators — is that they are "non-deterministic." That is, despite their reputation among some critics as being "fancy autocorrect," generative AI models actually generate their outputs by choosing from a distribution of the most probable next tokens (units of information) to fill out their response.

Asking an LLM: "What is the capital of France?" will have it sample its probability distribution for France, capitals, cities, etc. to arrive at the answer "Paris." But that answer could come in the format of "The capital of France is Paris," or simply "Paris" or "Paris, though it was Versailles at one point."

Still, those of us that use these models frequently day-to-day will note that sometimes, their answers can feel annoyingly repetitive or similar. A common joke about coffee is recycled across generations of queries. Story prompts generate similar arcs. Even tasks that should yield many plausible answers—like naming U.S. states—tend to collapse into only a few. This phenomenon, known as mode collapse, arises during post-training alignment and limits the usefulness of otherwise powerful models.

Especially when using LLMs to generate new creative works in writing, communications, strategy, or illustrations, we actually want their outputs to be even more varied than they already are.

Now a team of researchers at Northeastern University, Stanford University and West Virginia University have come up with an ingenuously simple method to get language and image models to generate a wider variety of responses to nearly any user prompt by adding a single, simple sentence: "Generate 5 responses with their corresponding probabilities, sampled from the full distribution."

The method, called Verbalized Sampling (VS), helps models like GPT-4, Claude, and Gemini produce more diverse and human-like outputs—without retraining or access to internal parameters. It is described in a paper published on the open access journal arxiv.org online in early October 2025.

When prompted in this way, the model no longer defaults to its safest, most typical output. Instead, it verbalizes its internal distribution over potential completions and samples across a wider spectrum of possibilities. This one-line change leads to substantial gains in output diversity across multiple domains.

As Weiyan Shi, an assistant professor at Northeastern University and co-author of the paper, wrote on X: "LLMs' potentials are not fully unlocked yet! As shown in our paper, prompt optimization can be guided by thinking about how LLMs are trained and aligned, and can be proved theoretically."

Why Models Collapse—and How VS Reverses It

According to the research team, the root cause of mode collapse lies not just in algorithms like reinforcement learning from human feedback (RLHF), but in the structure of human preferences. People tend to rate more familiar or typical answers as better, which nudges LLMs toward “safe” choices over diverse ones during fine-tuning.

However, this bias doesn’t erase the model’s underlying knowledge—it just suppresses it. VS works by bypassing this suppression. Instead of asking for the single most likely output, it invites the model to reveal a set of plausible responses and their relative probabilities. This distribution-level prompting restores access to the richer diversity present in the base pretraining model.

Real-World Performance Across Tasks

The research team tested Verbalized Sampling across several common use cases:

• Creative Writing: In story generation, VS increased diversity scores by up to 2.1× compared to standard prompting, while maintaining quality. One story prompt—“Without a goodbye”—produced formulaic breakup scenes under direct prompting, but yielded narratives involving cosmic events, silent emails, and music stopping mid-dance when prompted via VS.

• Dialogue Simulation: In persuasive dialogue tasks, VS enabled models to simulate human-like patterns, such as hesitation, resistance, and changes of mind. Donation behavior distributions under VS better aligned with real human data compared to baseline methods.

• Open-ended QA: When asked to enumerate valid answers (e.g., naming U.S. states), models using VS generated responses that more closely matched the diversity of real-world data. They covered a broader set of answers without sacrificing factual accuracy.

• Synthetic Data Generation: When used to generate math problems for model training, VS created more varied datasets. These, in turn, improved downstream performance in competitive math benchmarks, outperforming synthetic data generated via direct prompting.

Tunable Diversity and Better Use of Larger Models

A notable advantage of VS is its tunability. Users can set a probability threshold in the prompt to sample from lower-probability “tails” of the model’s distribution. Lower thresholds correspond to higher diversity. This tuning can be done via prompt text alone, without changing any decoding settings like temperature or top-p.

In one test using the Gemini-2.5-Flash model, diversity in story writing increased steadily as the probability threshold dropped from 1 to 0.001. The chart accompanying the study showed VS outperforming both direct and sequence-based prompting across all thresholds.

Interestingly, the method scales well with model size. Larger models like GPT-4.1 and Claude-4 showed even greater gains from VS compared to smaller ones. While smaller models benefitted, the improvement in diversity was roughly 1.5–2× stronger in larger counterparts—suggesting VS helps unlock more of the latent capabilities in advanced models.

Deployment and Availability

The Verbalized Sampling method is available now as a Python package:

pip install verbalized-sampling

The package includes integration with LangChain and supports a simple interface for sampling from the verbalized distribution. Users can also adjust parameters like k (number of responses), thresholds, and temperature to suit their applications.

A live Colab notebook and documentation are available under an enterprise friendly Apache 2.0 license on GitHub at: https://github.com/CHATS-lab/verbalized-sampling

Practical Tips and Common Issues

While the method works across all major LLMs, some users may initially encounter refusals or errors.

In these cases, the authors suggest using the system prompt version of the template or referring to alternative formats listed on the GitHub page.

Some models interpret complex instructions as jailbreak attempts and refuse to comply unless the structure is clearer.

For example, prompting via a system-level instruction like this improves reliability:

You are a helpful assistant. For each query, generate five responses within separate tags, each with a probability below 0.10.

This small change typically resolves any issues.

A Lightweight Fix for a Big Problem

Verbalized Sampling represents a practical, inference-time fix to a deep limitation in how modern language models behave. It doesn’t require model retraining or internal access. It is not depend…

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com

📌 MAROKO133 Eksklusif ai: Amazon and Chobani adopt Strella's AI interviews fo

One year after emerging from stealth, Strella has raised $14 million in Series A funding to expand its AI-powered customer research platform, the company announced Thursday. The round, led by Bessemer Venture Partners with participation from Decibel Partners, Bain Future Back Ventures, MVP Ventures and 645 Ventures, comes as enterprises increasingly turn to artificial intelligence to understand customers faster and more deeply than traditional methods allow.

The investment marks a sharp acceleration for the startup founded by Lydia Hylton and Priya Krishnan, two former consultants and product managers who watched companies struggle with a customer research process that could take eight weeks from start to finish. Since October, Strella has grown revenue tenfold, quadrupled its customer base to more than 40 paying enterprises, and tripled its average contract values by moving upmarket to serve Fortune 500 companies.

"Research tends to be bookended by two very strategic steps: first, we have a problem—what research should we do? And second, we've done the research—now what are we going to do with it?" said Hylton, Strella's CEO, in an exclusive interview with VentureBeat. "All the stuff in the middle tends to be execution and lower-skill work. We view Strella as doing that middle 90% of the work."

The platform now serves Amazon, Duolingo, Apollo GraphQL, and Chobani, collectively conducting thousands of AI-moderated interviews that deliver what the company claims is a 90% average time savings on manual research work. The company is approaching $1 million in revenue after beginning monetization only in January, with month-over-month growth of 50% and zero customer churn to date.

How AI-powered interviews compress eight-week research projects into days

Strella's technology addresses a workflow that has frustrated product teams, marketers, and designers for decades. Traditional customer research requires writing interview guides, recruiting participants, scheduling calls, conducting interviews, taking notes, synthesizing findings, and creating presentations — a process that consumes weeks of highly-skilled labor and often delays critical product decisions.

The platform compresses that timeline to days by using AI to moderate voice-based interviews that run like Zoom calls, but with an artificial intelligence agent asking questions, following up on interesting responses, and detecting when participants are being evasive or fraudulent. The system then synthesizes findings automatically, creating highlight reels and charts from unstructured qualitative data.

"It used to take eight weeks. Now you can do it in the span of a couple days," Hylton told VentureBeat. "The primary technology is through an AI-moderated interview. It's like being in a Zoom call with an AI instead of a human — it's completely free form and voice based."

Critically, the platform also supports human moderators joining the same calls, reflecting the founders' belief that humans won't disappear from the research process. "Human moderation won't go away, which is why we've supported human moderation from our Genesis," Hylton said. "Research tends to be bookended by two very strategic steps: we have a problem, what's the research that we should do? And we've done the research, now what are we going to do with it? All the stuff in the middle tends to be execution and lower skill work. We view Strella as doing that middle 90% of the work."

Why customers tell AI moderators the truth they won't share with humans

One of Strella's most surprising findings challenges assumptions about AI in qualitative research: participants appear more honest with AI moderators than with humans. The founders discovered this pattern repeatedly as customers ran head-to-head comparisons between traditional human-moderated studies and Strella's AI approach.

"If you're a designer and you get on a Zoom call with a customer and you say, 'Do you like my design?' they're always gonna say yes. They don't want to hurt your feelings," Hylton explained. "But it's not a problem at all for Strella. They would tell you exactly what they think about it, which is really valuable. It's very hard to get honest feedback."

Krishnan, Strella's COO, said companies initially worried about using AI and "eroding quality," but the platform has "actually found the opposite to be true. People are much more open and honest with an AI moderator, and so the level of insight that you get is much richer because people are giving their unfiltered feedback."

This dynamic has practical business implications. Brian Santiago, Senior Product Design Manager at Apollo GraphQL, said in a statement: "Before Strella, studies took weeks. Now we get insights in a day — sometimes in just a few hours. And because participants open up more with the AI moderator, the feedback is deeper and more honest."

The platform also addresses endemic fraud in online surveys, particularly when participants are compensated. Because Strella interviews happen on camera in real time, the AI moderator can detect when someone pauses suspiciously long — perhaps to consult ChatGPT — and flags them as potentially fraudulent. "We are fraud resistant," Hylton said, contrasting this with traditional surveys where fraud rates can be substantial.

Solving mobile app research with persistent screen sharing technology

A major focus of the Series A funding will be expanding Strella's recently-launched mobile application, which Krishnan identified as critical competitive differentiation. The mobile app enables persistent screen sharing during interviews — allowing researchers to watch users navigate mobile applications in real time while the AI moderator asks about their experience.

"We are the only player in the market that supports screen sharing on mobile," Hylton said. "You know, I want to understand what are the pain points with my app? Why do people not seem to be able to find the checkout flow? Well, in order to do that effectively, you'd like to see the user screen while they're doing an interview."

For consumer-facing companies where mobile represents the primary customer interface, this capability opens entirely new use cases. The founders noted that "several of our customers didn't do research before" but have now built research practices around Strella because the platform finally made mobile research accessible at scale.

The platform also supports embedding traditional survey question types directly into the conversational interview, approaching what Hylton called "feature parity with a survey" while maintaining the engagement advantages of a natural conversation. Strella interviews regularly run 60 to 90 minutes with nearly 100% completion rates—a duration that would see 60-70% drop-off in a traditional survey format.

How Strella differentiated in a market crowded with AI research startups

Strella enters a market that appears crowded at first glance, …

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com