Competitive Release: Testing Whether Language Models Hide Unstated Commonsense Behind Frequent Associations

Language models are bad at facts people rarely bother to say aloud. Examples are that glass is rigid, that a sheep is usually white, or that sleeping lasts hours rather than minutes. The standard explanation is that these facts are missing because text omits them, a pattern called reporting bias. This note tests a different explanation drawn from ecology. The facts may be present inside the model but crowded out at the point where the model chooses its answer, displaced by associations that text mentions often. Ecologists distinguish "this species is absent from the habitat" from "this species is present but suppressed by a dominant competitor" with a simple field experiment. They remove the dominant competitor and watch whether the subordinate expands. That expansion is called competitive release. We propose the same experiment inside a network. We locate the directions in the model's hidden activations that carry frequent text associations, remove them during the forward pass, and measure whether correct judgments about unstated properties come back.

The question matters because the two explanations call for different fixes. If the knowledge is absent, only new data or grounding helps. If it is suppressed, the fix is internal and cheap, and we can predict which facts are suppressed by measuring how much their internal representation overlaps with that of the dominant associations. Either answer is informative, and the experiment is designed so that both outcomes are clearly distinguishable.

Core Claim

Our hypothesis is that a measurable fraction of the commonsense failures attributed to reporting bias are failures of competitive exclusion at readout. Unstated perceptual and temporal properties are encoded at intermediate layers. They lose influence over the output in later layers to directions tied to high co-occurrence associations. Three predictions follow.

  1. Ablating the dominant association directions ("removal") should recover more accuracy than adding the property direction ("boosting").
  2. The size of the recovery should be predicted by a niche-overlap score between the property and its dominant competitor.
  3. Properties with no recoverable internal signal under any intervention should be classified as genuinely absent.

The resulting map of suppressed and absent properties across a commonsense benchmark is the main deliverable. A recovery method is secondary.

Four papers motivate the hypothesis.

  • Do Neural Language Representations Learn Physical Commonsense? found BERT strong on object–affordance pairs (0.88 micro-F1) and weak on affordance–property pairs (0.25 macro-F1 for properties versus 0.67 for humans). An example of the first kind is "car" with "drive." That paper used external probes and fine-tuning heads, so it cannot say whether the property signal is missing or merely unused.
  • Do Neural Language Models Overcome Reporting Bias? showed that models amplify textual skew beyond raw n-grams. Measured by KL divergence to real action frequencies, n-grams scored 2.94 and BERT and GPT-2 scored 3.77. That pattern looks like a dominant signal winning, not merely like data scarcity. The study worked entirely at the prompt level.
  • Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task supplies the method. Editing hidden activations to create counterfactual Othello boards changed moves accordingly. Our setting differs because it is open-world text, where the relevant facts are rarely stated.

What Ecology Contributes, and What It Does Not

We borrow three pieces of experimental design from ecology. We do not borrow its equations.

The first is the removal experiment, which separates absence from suppression. The second is resource-use overlap as a predictor of exclusion. The third is the observation that subordinates often survive by using a resource at a different time. In Water Use Patterns of Sympatric Przewalski's Horse and Khulan, dominant horses monopolize water points and khulans drink at off-peak hours. In Biodiversity is enhanced by sequential resource utilization and environmental fluctuations via emergent temporal niches, sequential use of resources lets far more species coexist than shared simultaneous use allows.

Our analogue of "time" is depth in the forward pass, meaning which layer a feature acts at. Our analogue of the contested resource is the subspace of the final hidden state that the output layer reads.

We make no claim that transformers obey consumer–resource dynamics. The ecology earns its place only if removal and overlap predict behavior better than the obvious alternatives below.

To make overlap concrete, let $e_f(\ell)$ be the causal effect on the correct-answer logit of patching feature $f$ at layer $\ell$, normalized to sum to one over layers. The niche overlap of property feature $p$ and dominant feature $d$ is

$$O(p,d) = \frac{\sum_\ell e_p(\ell)\, e_d(\ell)}{\sqrt{\sum_\ell e_p(\ell)^2 \sum_\ell e_d(\ell)^2}} \cdot \lvert \cos(u_p, u_d) \rvert,$$

where $u_p$ and $u_d$ are the two feature directions at the readout layer. The first factor measures overlap in depth and the second measures overlap in readout space. We predict that suppression, defined as the accuracy recovered by removing $d$, rises with $O(p,d)$.

Research Program

Models. We study open models that allow full activation access: BERT-large, RoBERTa-large, GPT-2 XL, and the Pythia suite. Pythia's training checkpoints let us watch features form over training. This connects to the finding in Bergen et al. that intermediate "quiet features" appear long before loss shows progress.

Finding the property feature. For each property (rigid, smooth, white, lasts-hours, and so on), we estimate a direction from items where the text states the property unambiguously. We use two methods:

A sparse autoencoder here is a model that rewrites hidden activations as a few active, nameable features. We use the causal-loss version because that paper shows ordinary sparse autoencoders often find features that correlate perfectly with a concept yet do nothing when clamped.

Finding the dominant feature. For each query, the dominant competitor is the direction whose activation tracks corpus co-occurrence strength between the object and candidate answers, using n-gram counts. It is estimated on "disagreement items" where co-occurrence and truth point opposite ways. Examples are "black sheep" phrasing versus white sheep, or murder versus breathing frequency.

Handling redundancy. Missed Causes and Ambiguous Effects shows that single-feature knockouts produce false negatives when backup features exist. For that reason, removal always operates on the full set of top dominant features at once. We also run denoising patches, which copy clean activations into a corrupted run rather than knocking activations out of a clean one.

Intervention conditions. We compare three conditions at matched intervention norm: removal (project out the dominant features), boosting (add the property direction), and both together.

Training-time analysis. Across Pythia checkpoints, we test a succession prediction. The property feature should have causal effect early in training, then lose its share of the readout as the dominant association strengthens.

Baselines

Do nothing. The naive reference is the unmodified model's zero-shot judgment. On the Situated dataset, the strongest reported number in the file is BERT's affordance–property macro-F1 of 0.25 for properties and 0.36 for affordances, against human scores of 0.67 and 0.65 (Do Neural Language Representations Learn Physical Commonsense?). Those BERT figures come from a trained head. We will rerun that head on our models with identical splits so that every comparison shares data.

Layer-wise probe. A non-linear probe is trained on the best intermediate layer, with the same labeled budget we use to estimate directions. This is the strongest competitor on the accuracy question. If information is merely present mid-network, a probe reads it out directly.

Fine-tuning and symbolic constraints. For color, the file reports BERT fine-tuned on truth data at about 70% versus about 50% zero-shot (Do Neural Language Models Overcome Reporting Bias?). For temporal duration, SLEER reached 83.0% on TimeML versus 73.7% for vanilla BERT by adding soft-logic rules to fine-tuning (Mitigating Reporting Bias in Semi-supervised Temporal Commonsense Inference with Probabilistic Soft Logic). On McTACO, the reported BERT with unit normalization scores 69.9% F1 and 42.7% exact match, against human 87.1% and 75.8% (Zhou et al.). These are references for what weight updates achieve. Our interventions change no weights.

Answer-prior calibration. This baseline subtracts each answer's average logit over content-free prompts. It tests whether "dominance" is just a global output bias rather than an input-specific internal competitor.

Key ablation. Boosting alone at matched norm isolates the competitive ingredient.

What counts as winning. A win requires two things. First, removal recovers significantly more accuracy on disagreement items than boosting and calibration, with confidence intervals bootstrapped over items. Second, the overlap score predicts per-property recovery with a rank correlation that a shuffled-pairing control does not reach. Whether removal beats the probe is reported but not required, because the probe answers a different question.

Why not just…?

Why not just probe the middle layers? This is the real threat on accuracy, and we concede it. A supervised probe may match or beat removal. Probing cannot say whether the model's own output uses the signal. Othello-GPT showed that decodability and causal control can come apart (Emergent World Representations), and causal-loss sparse autoencoder work found decodable features with zero downstream effect. Our claim is the diagnosis of suppressed versus absent, which only intervention provides. If the probe and removal recover the same properties, that agreement supports the suppression reading rather than replacing it.

Why not just steer toward the property? Boosting needs no ecological story. If boosting matches removal, the competition framing adds nothing, and we will report that. There is a reason to expect it to fall short. Endogenous Resistance to Activation Steering in Language Models shows models internally counteract added directions mid-generation, which is what a still-active dominant competitor would do. The matched-norm boost condition settles the question directly.

Why not just scale or fine-tune? The file reports two results against scale. Model size and corpus frequency do not repair affordance–property failures: there is no correlation with frequency in a 1-trillion-word corpus. Larger models also increased divergence from real action frequencies. Fine-tuning and soft-logic rules do help (83.0% on TimeML for SLEER). However, they change weights, need hand-written rules per property, and leave open whether the internal geometry changed. We will run our interventions on a SLEER-style fine-tuned model to test whether fine-tuning works by suppressing the dominant feature, which would be an informative result in itself.

Why not just recalibrate the answer prior? If dominance were only a fixed output bias, calibration would erase it. Our tasks use fixed yes/no answers across items, so a global prior shift cannot produce item-specific recovery. If calibration nonetheless matches removal, we will reframe the finding as an output-bias effect.

Proposed Experiments and Decisive Checks

Benchmarks. We use the Situated and Abstract datasets for physical properties (no reported number found on Lacuna for the Abstract dataset), Wikipedia color probing, McTACO for temporal properties, and the action-frequency KL task.

Disagreement items. On each benchmark, we construct the subset where n-gram co-occurrence favors the wrong answer. Suppression should appear there and nearly vanish on agreement items. If it appears on agreement items too, the intervention is simply adding noise.

First decisive check. Removal must beat boosting and calibration on disagreement items. If no intervention recovers accuracy for most properties, the absence explanation of reporting bias stands for those properties, and that map is the result.

Second check. Across properties, $O(p,d)$ should rank recovery above chance. Perceptual properties are predicted to show higher overlap and larger release than verb-like functional properties. This is consistent with Holtzman et al.'s finding that functional properties are learned better.

Third check. In Pythia checkpoints, the property feature's causal effect at readout should rise and then fall as the dominant feature grows. If instead it never forms, the failure is absence rather than exclusion.

Action-frequency check. Removal should reduce the KL divergence toward real frequencies, ideally below the 2.94 of raw n-grams. That would show the amplification above n-grams is internally reversible.