Why does combining natural language with numerical scores improve prediction accuracy?
This explores why pairing words with numbers — critiques alongside reward scores, narrative context alongside time-series values — tends to beat either signal alone, and the corpus traces it to a single idea: a number tells you *what*, but language tells you *why* and *in what context*.
This explores why pairing words with numbers tends to beat either alone — and the corpus keeps returning to one explanation: a scalar score is informationally thin. It can tell a model that it was wrong, but not why it was wrong or how to fix it. The clearest demonstration is in reinforcement learning, where models stall on plateaus that more numerical reward simply cannot break through. When the same models are handed chain-of-thought critiques instead of (or alongside) the score, they start producing correct solutions — because the language carries the missing diagnostic content about the *mechanism* of failure that a single number compresses away Can natural language feedback overcome numerical reward plateaus?.
The forecasting work reframes the same point as a division of labor. Numerical extrapolation and contextual reasoning are genuinely different cognitive jobs, and forcing one model to do both at once muddies both. When the architecture *separates* them — a stage for crunching the numbers, a stage for the event-driven narrative, then a synthesis — accuracy jumps past both pure time-series models and pure language models Can decomposing forecasting into stages unlock numerical and contextual reasoning?. The striking follow-on is that this isn't the model getting smarter; the forecasting ability was latent all along and only *surfaced* once the workflow stopped blending the two signals into one monolithic prompt Can LLMs actually forecast time series better than we think?. So combining the modalities helps partly because keeping them distinct lets each be done well before they're joined.
There's a deeper reason language is such a potent partner to numbers here: it's the substrate where a model integrates patterns across everything it knows. The same pattern-completion tendency that produces hallucination on backward-looking retrieval becomes genuine predictive power on forward-looking tasks — fine-tuned LLMs out-predict human neuroscientists on which experiments actually worked, precisely because language lets them fuse scattered contextual cues a bare number never encodes Can LLMs predict novel scientific results better than experts?. In the same spirit, models trained on psychology-experiment data beat hand-built theoretical models at predicting human choices, because the linguistic representation captures individual context that a clean numerical model throws away Can language models learn to model human decision making?.
But the corpus also adds an important caution, which is the thing you might not have known you wanted to know: the numbers aren't just decoration on the words either. A well-calibrated confidence score can itself be a *teaching* signal — using a model's own answer-span confidence to rank reasoning traces both sharpens reasoning and repairs the calibration that human-feedback training tends to wreck Can model confidence work as a reward signal for reasoning?. And a model that knows how to attach an honest number to its own uncertainty — when to abstain, when to retrieve — can match models ten times its size Can models learn to abstain when uncertain about predictions?, Can simple uncertainty estimates beat complex adaptive retrieval?. So the gain from combining isn't that language is rich and numbers are crude; it's that the two carry *non-overlapping* information — language carries causes and context, numbers carry magnitude and calibrated confidence — and prediction needs both.
Worth knowing where this can go wrong: language is a leaky channel. Models often track the *surface frequency* of phrasings rather than their meaning Do language models really understand meaning or just surface frequency?, and strong training-time associations can override the very context you fed them in Why do language models ignore information in their context?. That's exactly why grounding language in a numerical signal — a reward, a calibrated confidence, a verifiable score — helps: the number anchors the words to something the model can't fluently talk its way around.
Sources 10 notes
Critique-GRPO shows that models stuck on performance plateaus can generate correct solutions when given chain-of-thought critiques, revealing that numerical rewards lack critical information about why failures occur and how to improve.
Nexus outperforms pure TSFM and LLM baselines on real-world datasets by decomposing forecasting into contextualization, dual-resolution macro/micro outlook, and synthesis stages. Separating numerical extrapolation from event-driven contextual reasoning avoids forcing one model to handle both simultaneously.
LLMs have stronger intrinsic forecasting ability than recognized, but only when workflows separate numerical reasoning from contextual reasoning. Monolithic prompting obscures this capability; structured decomposition surfaces it.
BrainBench benchmarks show fine-tuned LLMs outperform neuroscience experts at predicting which experimental results actually occurred. The same pattern-integration tendency that causes hallucination in retrieval tasks enables genuine prediction in forward-looking scenarios.
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.
RLSF uses answer-span confidence to rank reasoning traces, creating synthetic preferences that strengthen step-by-step reasoning while reversing RLHF's calibration degradation—without requiring human labels or external verifiers.
Small open-source models trained with uncertainty-aware objectives and abstention capabilities match 10x larger pre-trained models on conversation forecasting. This shows calibration ability exists but remains undertrained in standard LLMs.
Calibrated token-probability uncertainty consistently beats multi-call adaptive retrieval on single-hop tasks and matches performance on multi-hop, using a fraction of the LM and retriever calls. The model's self-knowledge proves more reliable than external heuristics for deciding when to retrieve.
LLMs show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.