Abstract—Adaptive Inference Core (AIC) is a modular research framework for guiding language models during inference without corrupting the base model's parameters or original characteristics. Its premise is that models sharing similar architectures can display markedly different inference behavior, which makes parameter count or a single aggregate score an unreliable proxy for inference capability. AIC begins by characterizing a target model's behavioral profile, then uses a modular observation component to monitor its inference dynamics and feed the resulting signals into an adaptive state that guides subsequent generation. The framework operates as a closed loop: an Adaptive Guide conditions generation, a modular Inference Observer monitors the inference process, an Adaptive Engine evaluates what was observed, and a persistent Adaptive State carries that evaluation forward to the next cycle. Throughout, the base model remains the primary generator. The framework makes no assumption that hallucination or inference failure originates in any specific matrix, probability variable, or hidden representation; whether measurable inference dynamics can supply signals reliable enough to guide inference across both small and large language models remains the central open question this research is built to answer.
Index Terms—inference-time adaptation, behavioral profiling, closed-loop guidance, hallucination signals, cross-model generalization.
Modern language models generate output through a chain of transformations, from token embeddings through intermediate representations to output-token probabilities. A capable pretrained model does not guarantee that every inference trajectory draws on that knowledge equally well: the same model can be stable under one condition and inconsistent or poorly grounded under another.
AIC approaches this gap from an adaptive angle rather than a corrective one. Instead of asking how a model can be forced to produce a correct answer, it asks whether inference can be guided according to a model's observed behavioral characteristics and its present inference condition, without replacing the model's own reasoning or restricting the answers it is allowed to produce. Its purpose is narrower than an external answer generator: to guide inference adaptively while leaving the model's underlying capabilities intact.
AIC is organized as a closed adaptive loop of four modular components — the Adaptive Guide, the Inference Observer, the Adaptive Engine, and the Adaptive State — arranged around the target model's own inference pipeline, shown in Fig. 1.
$$\text{Observe} \rightarrow \text{Evaluate} \rightarrow \text{Update State} \rightarrow \text{Guide}$$
Unlike a conventional, static inference pipeline, information drawn from prior inference behavior shapes how AIC guides the inference that follows.
Adaptive inference presupposes some estimate of how the target model behaves. This should not collapse into a single function of scale, $\text{Capability} = f(\text{Params})$, nor a single scalar score. AIC instead adopts a multidimensional behavioral profile,
$$P_M = \{\, b_1, b_2, \ldots, b_n \,\}$$
where $P_M$ denotes the behavioral profile of model $M$ and each $b_i$ describes one observed property — candidates include contextual retention, response consistency, topic transition, instruction handling, and behavior across context lengths. It follows that
$$\text{Model Size} \not\Rightarrow \text{Inference Stability}$$
A small model should not be presumed unstable on that basis alone, nor a large model presumed stable — behavior has to be measured directly.
Given input $x_t$ at step $t$ and adaptive state $S_t$, the Adaptive Guide is represented as
$$G_t = \mathcal{G}(x_t, S_t, P_M)$$
where $G_t$ is the resulting guidance condition. $G_t$ does not encode a predetermined answer: AIC guides the conditions under which the model generates using its own capabilities, keeping the base model the primary generator throughout.
The Inference Observer is a modular component that monitors the inference process as it unfolds, without independently judging the semantic correctness of the output. Given intermediate representations $H^{(1)}, \ldots, H^{(L)}$, it is represented as
$$O_t = \mathcal{O}\left(H^{(1)}, \ldots, H^{(L)}\right)$$
This does not imply that any particular hidden-state pattern causes hallucination. At the current stage, the relevant question is narrower: whether measurable signals carry information useful for adaptive guidance at all. What signals to track, and how to extract them, remains an open modular design question.
Following generation, the Adaptive Engine processes what was observed and updates the state:
$$S_{t+1} = \mathcal{A}(S_t, P_M, O_t, Y_t)$$
giving the feedback loop $S_t \rightarrow G_t \rightarrow M \rightarrow O_t \rightarrow Y_t \rightarrow S_{t+1}$. The updated state becomes available to the Guide on the next cycle, so adaptation accumulates without redefining knowledge already in the base model.
Where a conventional pipeline treats every request independently ($x_1 \rightarrow y_1$, $x_2 \rightarrow y_2$), AIC introduces state continuity:
$$(x_t, S_t) \rightarrow y_t \rightarrow S_{t+1}$$
The Adaptive State is not a substitute for model weights: the base model's parameters $\theta_M$ remain distinct from the adaptive condition $S_t$, so AIC investigates adaptation at inference time while leaving the original model untouched.
AIC draws a firm line between observation, association, and causation. For layer $l$, a representation $H_l \in \mathbb{R}^{T \times d}$ changes across layers in a way that could be tracked by a general measurement function $D_l = \mathcal{D}(H_l, H_{l-1})$, yielding a trajectory $D_1, \ldots, D_L$. AIC makes no causal claim of the form $D_l\!\uparrow \Rightarrow \text{hallucination}$. Even if a future measurement $z$ correlates with hallucinated responses, this supports only
$$P(\text{failure} \mid z) \neq P(\text{failure})$$
under tested conditions — not $z \rightarrow \text{failure}$. Causality, if it exists, requires considerably stronger evidence than correlation.
This caution extends across models. A signal observed in one model cannot be generalized automatically to all Transformer-based language models — validation is required across both small models ($\text{SLM}_1, \ldots, \text{SLM}_n$) and large models ($\text{LLM}_1, \ldots, \text{LLM}_m$), testing whether a candidate signal is model-specific, scale-dependent, or general enough to serve as an adaptive indicator.
Taken together: language-model inference may contain measurable signals that an external, modular system can use for model-specific guidance without directly modifying the underlying model. Which variable, if any, carries the decisive signal is left open as an empirical question.
AIC remains an active research project. This work does not claim to have identified the universal cause of hallucination, nor that any particular matrix or probability variable causes inference failure. Findings from individual models should not be generalized to all SLMs and LLMs without cross-model validation. These are open empirical questions, not implementation defects.
AIC proposes a closed-loop, modular approach to inference-time adaptation: a target model is characterized behaviorally, its inference observed, and the resulting observations shape subsequent guidance through a persistent adaptive state,
$$\text{Profile} \rightarrow \text{Guide} \rightarrow \text{Infer} \rightarrow \text{Observe} \rightarrow \text{Adapt} \rightarrow \text{State}$$
Its aim is not to replace a model's reasoning or dictate predetermined answers, but to test whether a modular, external system can observe a model, understand its behavioral tendencies, and guide it toward more stable use of the capabilities it already has.