Operationalising the Superficial Alignment Hypothesis via Task Complexity

Abstract: The superficial alignment hypothesis (SAH) posits that LLMs learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge. The SAH, however, lacks a precise definition, which has led to (i) different and seemingly orthogonal arguments supporting it, and (ii) important critiques to it. We propose a new metric called task complexity: the length of the shortest program that achieves a target performance on a task. In this framework, the SAH simply claims that pre-trained models drastically reduce the complexity of achieving high performance on many tasks. Our definition unifies prior arguments supporting the SAH, interpreting them as different strategies to find such short programs. Experimentally, we estimate the task complexity of mathematical reasoning, machine translation, and instruction following; we then show that these complexities can be remarkably low when conditioned on a pre-trained model. Further, we find that pre-training enables access to strong performances on our tasks, but it can require programs of gigabytes of length to access them. Post-training, on the other hand, collapses the complexity of reaching this same performance by several orders of magnitude. Overall, our results highlight that task adaptation often requires surprisingly little information -- often just a few kilobytes.

• The paper introduces a quantitative measure using the shortest adaptation program length to evaluate LLM adaptation efficiency.
• It demonstrates that diverse adaptation protocols—data-driven fine-tuning, parameter-efficient updates, and inference-time prompting—populate distinct regions on a Pareto frontier.
• The study empirically shows that post-training significantly reduces adaptation complexity, enabling strong performance with minimal additional information.

A Formal Operationalization of the Superficial Alignment Hypothesis via Task Complexity

Introduction and Context

The superficial alignment hypothesis (SAH) asserts that LLMs acquire the vast majority of their knowledge during pre-training, while downstream adaptation procedures – including data-based fine-tuning, parameter-efficient transfer, and inference-time prompting – act primarily as mechanisms for revealing or "surfacing" this already-encoded knowledge. Although SAH has generated significant interest, prior discourse has been hampered by vague definitions of key terms such as “knowledge”, “capabilities”, and “formats”. As a result, evidence for and against SAH has been drawn from seemingly orthogonal perspectives, leading to intellectual fragmentation and robust critiques.

This work provides a rigorous operationalization of SAH, grounded in algorithmic information theory, by defining the notion of task complexity and leveraging it to unify disparate supporting arguments. The central proposal is to quantify the amount of information required to adapt an LLM to a target performance on a task in terms of the shortest program ("adaptation script") that leverages the pre-trained model’s weights. This metric enables direct, quantitative comparison of “superficiality” across different adaptation modalities.

Task Complexity: A Unified Metric for Adaptation

The authors define a task $\mathcal{T}$ as consisting of input and output spaces, a distribution over inputs, and a scoring function for evaluating outputs. The complexity of achieving a target performance $\delta$ on $\mathcal{T}$ is formalized as the length (in bits) of the shortest program that, possibly making use of a pre-trained model $\mathcal{M}$, achieves expected performance at least $\delta$. The conditional task complexity $\mathcal{C}({\mathcal{T}_\delta} \mid \mathcal{M})$ thereby quantifies how much information, in addition to the weights of $\mathcal{M}$, is needed to accomplish $\mathcal{T}$ up to accuracy $\delta$.

The central claim of the SAH, then, becomes: for many complex tasks of practical interest, $\mathcal{C}({\mathcal{T}_\delta} \mid \mathcal{M})$ is small when $\mathcal{M}$ is a pre-trained LLM.

Three previously “orthogonal” literatures are interpreted as proposing distinct strategies for constructing short programs:

• Data View: Minimal adaptation data (few-shot or small fine-tuning sets) suffice, focusing on the description length of training data.
• Parametric View: Only a small number of weights (or low-rank adapters) need be updated, focusing on the length of the parameter delta.
• Inference-Control View: Frozen models can be adapted at inference via compressed prompts, with program length corresponding to prompt encoding.

Critically, programs arising from each view are placed on a length–performance Pareto curve, enabling comparison not only across methods, but also across models and tasks.

Figure 1: The Pareto curve of program length vs. performance for Olmo3-7B on GSM8K; distinct views on superficiality map to different regions of this trade-off.

Experimental Design

To instantiate the operationalization, the authors consider three tasks:

• Mathematical reasoning (GSM8K/MetaMath)
• Machine translation (FLORES-200, En-Fr)
• Instruction following (IFEval)

They employ publicly available models: SmolLM3 (3B), Olmo3 (7B and 32B), and multiple adaptation modalities (full fine-tuning, parameter-efficient methods such as LoRA and Bayesian-LoRA, standard and instruction-augmented in-context learning). For each, they measure the total message length required for adaptation, encompassing not only data or parameters, but also required scripts (under the assumption of Python as the UTM).

Pareto curves are constructed by exhaustive hyperparameter sweeps, admitting only non-dominated protocol configurations, and program lengths are referenced against familiar quantities (e.g., single ImageNet image size) to calibrate intuition.

Figure 3: Program length vs. performance Pareto curves for SmolLM3 and Olmo3 across three tasks, displaying optimal adaptation strategies.

Main Empirical Findings

1. Programs of Kilobyte to Megabyte Lengths Achieve Strong Adaptation:
   • High performance on complex tasks can often be achieved by encoding adaptation programs of just a few kilobytes to (at most) several megabytes, despite model pre-training involving terabytes of data.
   • For Olmo3-32B on GSM8K, a program of $\approx$4,358 bits enables 72.2% accuracy; for Olmo3-7B on FLORES, programs under 4,000 bits suffice for BLEU of 34.43.
   • The efficiency gap between adaptation and pre-training highlights the superficiality of post-training stages.
2. Different Adaptation Protocols Populate Different Regions of the Pareto Frontier:
   • Inference-control (prompt-based) methods yield the shortest programs with modest accuracy improvements.
   • Data-driven fine-tuning occupies the mid-range, delivering strong accuracy at moderate program lengths.
   • Parameter-efficient methods (LoRA, Bayesian-LoRA) are effective in the high-length, maximum-performance regime.

     Figure 2: The Pareto curves for SmolLM3 3B, Olmo3 7B, and Olmo3 32B, with distinct adaptation methods contributing complementary regions.

3. Pre-Training Enables, But Does Not Always Surface, High Performance:
   • Pre-trained models permit access to strong performance, but in some cases only via extremely long programs (gigabytes), indicating hidden capacity.
   • Post-training dramatically collapses the required adaptation complexity: after post-training, high accuracy can be obtained with vastly smaller programs, confirming the “surfacing” effect.

     Figure 6: Estimated complexity curves conditioned on pre-trained and post-trained checkpoints for SmolLM3 and Olmo3; post-training reduces program size required for strong performance by orders of magnitude.

Analysis of Prior Critiques/Claims

The information-theoretic framework allows formal reassessment of critiques of SAH:

• Performance saturation with few-shot data is not always possible: High accuracy may require much information even if moderate accuracy is accessible with very short programs. This nuance resolves disputes arising from ambiguous definitions of "knowledge" and "capability."
• Delta compression estimates can be loose: Methods that compress parameter updates to, e.g., 1 bit per parameter, generally produce programs far from the Pareto-optimal message length. Data-driven adaptation may require much less information.
• Head-only fine-tuning is suboptimal: Restricting to linear output layer adaptation under the guise of “superficial knowledge” may select suboptimal programs; this view is not justified as unique for operationalizing SAH, as shown by poor performance–length trade-off compared to more flexible methods.

  Figure 5: Direct comparison of linear projection fine-tuning and Pareto-efficient adaptation routes for Olmo3-7B on GSM8K.

Practical and Theoretical Implications

This operationalization of SAH exposes adaptation as fundamentally an information compression problem, where the shortest adaptation programs quantifiably demonstrate the “superficiality” of downstream learning. The implications include:

• Paradigm unification: Prior arguments about prompt engineering, parameter-efficient fine-tuning, and small-data transfer are not contradictory but represent search strategies for short, performant programs.
• Algorithmic diagnostics: The shape of the Pareto frontier for a given model and task gives insight into where adaptation becomes bottlenecked—by information, search, or architecture.
• Model release safety: The complexity of extracting new (possibly dangerous) capabilities from open-weight models could, in principle, be measured in this framework by estimating the minimal information programs required for task instantiation.
• Limits and estimation gap: Although upper bounds on task complexity are computable via search, the true minima are generally intractable to find. As a result, all empirical results are upper bounds, but systematic exploration of adaptation protocols is an important research direction.

Conclusion

The authors present a theoretically grounded, empirically supported formalism that enables precise discourse about superficial alignment in LLMs. By defining and estimating conditional task complexity, they unify various adaptation strategies as concrete instantiations of the search for short, high-performance programs. The work demonstrates that, for diverse and challenging NLP tasks, pre-trained LLMs can indeed be adapted with surprisingly little additional information. Post-training has the dual effect of increasing both peak attainable performance and—more importantly—collapsing the length of adaptation programs required.

The developed framework is immediately applicable to further inquiries into adaptation, safety, and compression in LLMs. Future research will be required to develop even more efficient program synthesis methods and to generalize the task complexity framework to richer, multi-modal, or open-ended problem settings.