Can a single model trained on two tasks predict untrained decision tasks?
This explores whether a model trained on a handful of tasks can generalize to decision tasks it never saw during training — the corpus speaks to this most directly through cognitive modeling, but also through what's known about where multi-task transfer comes from.
This explores whether a model trained on just a couple of tasks can predict decision tasks it was never trained on. The closest answer in the collection comes from work where language models were fine-tuned on psychology experiment data and turned into general-purpose models of human choice Can language models learn to model human decision making?. The striking result there is that such models predict human behavior in decision tasks more accurately than hand-built cognitive theories, capture differences between individual people in their internal representations, and — most relevant to your question — transfer across tasks without anyone designing a task-specific architecture. So the short version: yes, transfer to unseen decision tasks is observed, and it doesn't require building a custom model per task.
The more interesting question is *why* this works, and here the corpus pushes back on the intuition that training 'creates' the new ability. Several independent lines of evidence suggest the capability is already latent in the base model and training merely elicits it Do base models already contain hidden reasoning ability?. A related finding sharpens this: instruction tuning often teaches the model the *shape* of valid answers (the output space) rather than genuine task understanding — models trained on semantically empty or even wrong instructions perform about as well as those trained correctly Does instruction tuning teach task understanding or output format?. Read together, these suggest that 'training on two tasks then predicting a third' may be less about learning two tasks and more about the model recognizing which already-present competence to apply.
There's also a mechanism for how a model can hold many tasks at once. LLMs appear to represent multiple distinct tasks simultaneously in superposition during inference Can LLMs handle multiple tasks at once during inference? — though there's a catch worth knowing: autoregressive decoding collapses that superposition to a single task after the first token, which limits doing several at once in one generation. And for sequential decision-making specifically, generalizing to brand-new tasks without weight updates turns out to depend on the *structure* of what's in context — full or partial trajectories from the same environment, not isolated examples Why do trajectories matter more than individual examples for in-context learning?. That's a quiet but important caveat: untrained-task generalization in decision settings can hinge on giving the model trajectory-shaped context, not just examples.
If you want the architectural angle on combining tasks in one model, two notes are worth a look. Wide & Deep shows how a single model can fuse memorization and generalization by training both jointly, with each half specializing so the whole outperforms either alone Can one model memorize and generalize better than two?. And Transformer² composes task-specific 'expert vectors' at inference time, dynamically mixing skills for a new task without retraining Can models dynamically activate expert skills at inference time? — arguably the cleanest demonstration of 'recombine known pieces to handle something unseen.' The thread connecting all of this: predicting an untrained task is rarely about the two training tasks themselves and mostly about whether the relevant capability was already there to be selected, composed, or elicited.
Sources 7 notes
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Models trained on semantically empty or deliberately incorrect instructions achieve comparable performance to those trained on full correct instructions, achieving 43% vs random baseline 42.6%. The semantic content of instructions appears largely irrelevant; what transfers is knowledge of the output space.
Large language models represent multiple complete, computationally distinct tasks simultaneously during inference—a macroscopic phenomenon separate from feature-level superposition. However, autoregressive decoding forces convergence to a single task after the first token, preventing practical multi-task generation.
In-context learning for sequential decision-making requires full or partial trajectories from the same environment level, not just isolated examples. This structural property—trajectory burstiness—allows models to generalize across vastly different tasks without weight updates.
Wide & Deep models train memorization (cross-product features) and generalization (embeddings) together, allowing each component to specialize: the wide part becomes small because deep handles common cases, and deep doesn't overfit rare items because wide captures them. Ensembling requires both halves full-size.
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.