How LLMs can replicate human purchase intent

Recent advances in language models have demonstrated remarkable capabilities in text generation and contextual understanding, yet these models remain fundamentally static, once trained their weights no longer change. This limitation prevents models from adapting to new knowledge, specific contexts, or emerging tasks, reducing their ability to generalize and stay up-to-date. In the paper “Self-Adapting Language Models (SEAL)", the authors propose a framework that enables models to self-adapt autonomously, without requiring costly external fine-tuning or additional human annotations. The method relies on generating self-edits, which are revised or enriched versions of the model’s own outputs, serving as synthetic data for internal supervised fine-tuning. For each input, the model first produces an initial response and then generates a self-edit that can correct errors, rephrase the output, enrich content, or adjust training parameters. These self-edits are evaluated using task-specific metrics, and only those deemed effective are incorporated into the model’s weights, enabling persistent updates. This approach establishes a loop of continuous self-improvement, allowing the model to generate progressively more accurate and contextually appropriate responses while retaining prior knowledge. #SelfLearningAI #FewShotLearning #SyntheticData #ReinforcementLearning

Meta’s REFRAG: rethinking RAG decoding Just read Meta Superintelligence Labs’ new paper on REFRAG — a smart decoding framework for Retrieval-Augmented Generation that compresses retrieved passages, senses which chunks matter, and expands only the important bits during generation. The result is striking: up to 30.85× faster time-to-first-token, roughly 16× larger effective context, and lower token usage — without loss of accuracy across standard RAG benchmarks. Why this matters: real-world RAG systems (search+LLM) often jam huge retrieved contexts into the model, paying heavy latency and memory costs for mostly irrelevant tokens. REFRAG changes the execution pattern: do cheap, compressed checks for most chunks and only expand the handful that actually contribute — the win is both speed and scale. Practical implications: faster conversational agents, cost-efficient enterprise RAG at scale, and the possibility to run deep long-document reasoning with smaller LLMs. Meta will release code (and community ports already exist) — worth experimenting with for any production RAG pipeline. #RAG #LLM #AI #ML #NLP #AIEngineering #MetaResearch source : https://lnkd.in/gPuBpvES

I’d like to share two recent applied research papers ( published on September and October 2025) that caught my attention for their innovative contributions and potential to advance the industrialization of Generative AI — across both Large Language Models (LLMs) and Diffusion Models (a key branch of Generative AI focused on generating images, videos, and other modalities). 1️⃣ Meta AI — “REFRAG: Rethinking RAG-based Decoding” The authors propose REFRAG (REpresentation For RAG), a new mechanism for more efficient context decoding in Retrieval-Augmented Generation (RAG), significantly reducing inference latency. -paper link : https://lnkd.in/eat55KMB - open source code not yet published 2️⃣ New York University — “Diffusion Transformers with Representation Autoencoders” This paper introduces Representation Autoencoders (RAEs), which reuse pretrained, frozen encoders with lightweight trained decoders to provide high-fidelity, semantically rich latent representations for diffusion transformers. 📄 Learn more and access the official paper, code, and models here: https://rae-dit.github.io/ All credits to the original authors. Shared for educational and research discussion purposes.

🧑🏫 AI for product development: We've been doing it wrong. Two new research papers show exactly how to fix it - one from Wharton, one from PyMC Labs + Colgate-Palmolive. PAPER 1: Ideation (Wharton) When you ask GPT "generate 100 ideas," it produces ideas 50-75% MORE similar than human teams. Solution: Chain-of-Thought prompting → Generate titles only → Force divergence ("make bolder, more different") → Then full descriptions Result: 27% more unique ideas PAPER 2: Consumer Testing (PyMC Labs + Colgate) Traditional surveys: £10K, 2 weeks per concept Direct LLM ratings: Useless narrow distributions Solution: Semantic Similarity Rating (SSR) → Synthetic consumer personas → Natural text responses → Convert to ratings via embeddings Result: 90% accuracy vs humans, £50 vs £10K Tested on 9,300 real consumers across 57 products THE PATTERN: Both show: naive prompting fails, structured prompting transforms results. ❌ "Give me ideas" or "Rate this 1-5" ✅ Multi-step processes with explicit constraints THE CAVEATS: - CoT needs human filtering for feasibility - SSR works for consumer products, not B2B/technical   - Always validate finalists with real humans - Both require structured thinking and domain expertise THE OPPORTUNITY: Most companies are still "trying ChatGPT for everything." Winners will systematise AI for specific high-value use cases. These papers provide validated, immediately implementable approaches. Papers: 1. "Prompting Diverse Ideas" https://lnkd.in/egRS3u5Q 2. "LLMs Reproduce Human Purchase Intent"  https://lnkd.in/eY7paNaC For product leaders spending £50K+ on customer R&D: this deserves a pilot. Thoughts? Who's implementing this? 👇 #ProductDevelopment #AI #Innovation #MarketResearch

The Next Step in Self-Improving AI Systems: Agentic Context Engineering (ACE) Paper from  Qizheng Zhang. Researchers from Stanford University and SambaNova Systems have introduced ACE (Agentic Context Engineering) — a new framework that enables large language models (LLMs) to self-improve through evolving contexts, instead of retraining model weights. 🔍 Why it matters Current LLMs rely heavily on static prompts or fine-tuning. However, these methods often suffer from: Brevity bias – simplifying prompts too much and losing critical domain insights. Context collapse – rewriting prompts repeatedly until important details disappear. 💡 How ACE works ACE treats contexts as living playbooks that grow and refine themselves via three modular agents: 1. Generator – produces reasoning trajectories. 2. Reflector – analyzes successes and failures to extract insights. 3. Curator – updates the playbook incrementally. This approach preserves knowledge over time and adapts efficiently, even without labeled data. 📊 Key results +10.6% improvement on agent tasks and +8.6% on financial reasoning benchmarks. 86.9% lower adaptation latency and reduced rollout costs. Outperformed or matched GPT-4–based systems on the AppWorld leaderboard, using a smaller open-source model (DeepSeek-V3.1). 🌱 Why it’s a game changer ACE shows that context—not weights—can drive AI evolution. By continuously learning from feedback, LLMs can become more accurate, explainable, and efficient — paving the way for truly self-improving AI agents. #AI #LLM #Research #MachineLearning #AgenticAI #Stanford #SambaNova #ContextEngineering https://lnkd.in/d-gVGtit

🚀 The Next Leap for LLMs: From Static Models to Self-Improving Systems Two recent papers caught my attention this week — both pointing toward a future where language models don’t just generate knowledge, but evolve it. 🧩 1️⃣ Self-Adapting Language Models (MIT, NeurIPS 2025) https://lnkd.in/gtEkJhKs Introduces SEAL, a framework where an LLM can finetune itself — generating its own training data, defining its own optimization steps, and updating its own weights through reinforcement learning. It’s like giving the model not just a “brain,” but the ability to rewrite its own synapses. 🧠 2️⃣ Agentic Context Engineering (Stanford, SambaNova, UC Berkeley) https://lnkd.in/gP427FJR Proposes ACE, a system where an LLM continuously improves its prompts, memories, and reasoning traces — learning from experience without changing weights. Think of it as a model that keeps rewriting its notebook of strategies to become more effective over time. What’s fascinating is how these two directions complement each other: 💡 SEAL → focuses on internal self-modification (weights, finetuning). 💡 ACE → focuses on external self-modification (context, prompts, memory). Together, they sketch a vision of agentic, self-improving AI — systems that refine both what they know and how they think, autonomously. #AI #MachineLearning #LLM #NeurIPS2025 #ArtificialIntelligence #SelfImprovingAI #ReinforcementLearning #PromptEngineering

This paper shows that you can predict actual purchase intent (90% accuracy) by asking an LLM to impersonate a customer with a demographic profile, giving it a product & having it give its impressions, which another AI rates. No fine-tuning or training & beats classic ML methods. https://lnkd.in/gFY29x2J

Retrieval augmented generation (RAG) enhances large language models (LLMs) by providing them with relevant external context. For example, when using a RAG system for a question-answer (QA) task, the LLM receives a context that may be a combination of...

LLMs don’t make decisions—they evolve them. New paper shows you can get human-shaped purchase intent by prompting text first, then mapping it to Likert via semantic similarity. Translation: decisions live in a probability field, not a single number. That’s the Q2E thesis—recursive, distributional feedback beats brittle scoring. Why I care: this is real-world validation for quotient dynamics in market cognition. If your “AI insight” is just a 1–5, you’re flying blind. Use the field. Read: https://lnkd.in/g3tP_Rjb #AGI #Q2E #DecisionScience #LLM #MarketingScience #QuotientIntelligent

Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks (Beijing Jiaotong University, October 2025) Paper: https://lnkd.in/eu9YifRh Abstract: "Large Language Models face challenges in long-horizon agentic tasks as their constrained memory is easily overwhelmed by distracting or irrelevant context. Existing working memory methods typically rely on external, heuristic mechanisms that are decoupled from the agent's core policy. In this work, we reframe working memory management as a learnable, intrinsic capability. We propose a novel framework, Memory-as-Action, where an agent actively manages its working memory by executing explicit editing operations as part of a unified policy. This formulation allows an agent, trained via reinforcement learning, to balance memory curation against long-term task objectives under given resource constraints. However, such memory editing actions break the standard assumption of a continuously growing prefix in LLM interactions, leading to what we call trajectory fractures. These non-prefix changes disrupt the causal continuity required by standard policy gradient methods, making those methods inapplicable. To address this, we propose a new algorithm, Dynamic Context Policy Optimization, which enables stable end-to-end reinforcement learning by segmenting trajectories at memory action points and applying trajectory-level advantages to the resulting action segments. Our results demonstrate that jointly optimizing for task reasoning and memory management in an end-to-end fashion not only reduces overall computational consumption but also improves task performance, driven by adaptive context curation strategies tailored to the model's intrinsic capabilities."