What non-parametric methods could replace latent factors for inductive learning?
This reads 'latent factors' as the learned, parameter-baked representations that classic models rely on, and asks which memory- or retrieval-based (non-parametric) alternatives let a system generalize to new cases without retraining those parameters.
This explores what could stand in for latent factors — the dense learned representations baked into a model's weights — when the goal is inductive learning, i.e. generalizing to new tasks or entities without going back and retraining. The corpus's clearest answer is a family of memory-based methods that move the 'learning' out of the parameters and into an external, queryable store. Instead of compressing experience into latent factors, these systems keep experience around and reason over it at inference time.
The sharpest example is AgentFly, which reframes agent learning as a memory-augmented decision process: it carries case memory, subtask memory, and tool memory, and improves its policy entirely through memory operations — no weight updates at all — yet still hits 87.88% on the GAIA benchmark Can agents learn continuously from experience without updating weights?. Reflexion makes the same bet more minimally: an agent takes a binary success/failure signal, writes a verbal self-diagnosis, and stores it as episodic memory it can consult next episode Can agents learn from failure without updating their weights?. Both are non-parametric in the truest sense — the knowledge lives in retrievable records, not in latent coordinates — and both are inductive: they transfer to situations never seen during any training run. Notably, Reflexion's authors find that keeping reflections uncompressed matters; the moment you compress memory back toward a fixed-size representation, you start to lose the very thing that made it non-parametric.
Why reach for this at all? Partly because the corpus also documents where latent-space machinery quietly fails. LLMs asked to run iterative numerical methods 'in their heads' don't actually execute the procedure — they pattern-match memorized templates and emit plausible-but-wrong values, a failure that persists across scale Do large language models actually perform iterative optimization?. And when reasoning is decoupled from familiar semantics, model performance collapses even with correct rules in hand, because the work is being done by parametric association rather than manipulable structure Do large language models reason symbolically or semantically?. These are arguments for keeping knowledge external and inspectable rather than dissolved into latent factors.
The interesting tension is that not every alternative abandons latent space — some make it richer. Latent-Thought models add a fast 'local' learning loop over latent vectors that scales independently of parameters Can latent thought vectors scale language models beyond parameters?, and GRAM makes latent transitions stochastic so a reasoner can hold a distribution over solutions instead of one point Can stochastic latent reasoning help models explore multiple solutions?. Read against the memory methods, these mark the real fork in the road: you can either keep latent factors and give them more expressive dynamics, or you can replace them with an episodic store you can read, edit, and grow.
The thing worth taking away: the corpus's non-parametric substitutes aren't just 'a database bolted on.' They relocate credit assignment and policy improvement into memory operations — which means a system can keep learning after deployment, with frozen weights, simply by remembering better. That's a different shape of learning than tuning latent factors, and it's where several of these papers are quietly converging.
Sources 6 notes
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
Latent-Thought Language Models achieve superior sample and parameter efficiency by coupling fast local variational learning with slow global decoder learning. This dual-rate scheme scales few-shot reasoning across both model and latent size, creating independent scaling dimensions beyond traditional parameter scaling.
GRAM replaces deterministic latent updates with stochastic sampling, enabling models to represent distributions over solutions rather than single predictions. This allows handling of ambiguous problems and multiple valid strategies that deterministic designs cannot represent.