When people say someone is "not tall", they usually leave room for "average height". When they say "not very tall", they leave even more room. Language models may not leave this room. Recent mechanistic work found that models handle "not X" by building a replacement concept rather than removing X. In mechanistic work, "building" means the middle layers actively promote an alternative ("not gas" becomes "solid or liquid"). For an on/off property, that is a sensible strategy. For a property that comes in degrees, such as height, weight, or warmth, the natural replacement is the antonym ("short"), and the middle of the scale disappears. We propose to test this "excluded-middle" hypothesis directly. We will measure what value a model implies when it hears a negated vague description, compare it to human judgments, and locate where in the network the graded meaning is kept or lost.
The opportunity is that degree words are where three known weaknesses meet. Negation is fragile. Vague words lack crisp boundaries. Real-world scales such as height are entangled with memorized web statistics, so a model can look correct by recalling typical values rather than computing what the sentence means. We resolve the third problem by defining fictional scales inside the prompt, which gives a clean ground truth that pretraining could not have supplied. We then compare against familiar real scales to see how memorized typicality changes the internal computation.
Core Claim
Our claim is that, for vague scalar predicates, the internal mechanism of negation is built for antonyms and fits graded meaning poorly. Concretely, "not P" and "not very P" push the model's internal estimate of the underlying quantity past the midpoint toward the opposite pole. As a result, models put too little probability on the middle region that humans include. We further predict that this distortion is partly a late-layer effect. A linear readout of the quantity in middle layers should be closer to human judgments than the model's final output, mirroring the finding of How Language Models Process Negation that correct negation logic computed mid-network is overridden by late shortcut attention heads. If the middle-layer readout already shows the collapse, the problem lies in how the model represents negation itself, not in a late override. That outcome is equally informative, and it changes which repairs make sense.
Where This Sits
The proposal starts from How Language Models Process Negation. That paper studied 648 crisp prompts of the form "X that is not Y is Z" in Llama-3.1-8B and Mistral-7B. It found construction in layers 14–18 in over 80% of cases, against direct suppression in about 30%. It also showed that ablating late shortcut heads lifts negated-prompt accuracy from 50.5% to 67.8%. The paper explicitly leaves graded and hedged negation open. Its main readout, mapping intermediate states onto vocabulary tokens ("logit lens"), can only show which words are promoted. It cannot show a position on a continuous scale. We replace that token readout with a continuous magnitude probe.
That probe is possible because (Language Models Encode the Value of Numbers Linearly; Language Models Encode the Value of Numbers Linearly) showed that numerical magnitude is linearly encoded in hidden states and causally used. We borrow the method. We do not borrow the claim that magnitude probes transfer across contexts: Investigating Linear Probe Robustness to Linguistic Register, Medical Specialty, and Corpus Shifts in Medical QA warns that probes often fail under shift, so probe transfer from affirmative to negated sentences must be validated rather than assumed.
The normative side comes from fuzzy semantics. In this tradition a word like "tall" is assigned a degree of membership between 0 and 1 rather than true or false (PRUF). Hedges like "very" act as operators that sharpen membership, as formalized in the Khayyam language of Lexical Imprecision in Fuzzy Constraint Networks. We use the textbook calculus, in which "not" is $1-\mu$ and "very" is $\mu^2$, as one reference prediction. Human judgments remain the primary target, because Colour Me Uncertain found that graded probabilistic semantics beat fuzzy t-norm listeners on human color reference (88.18% vs 87.96%). Fuzzy logic is therefore a hypothesis about people, not ground truth.
Two behavioral studies motivate the work and also constrain it. Probing Quantifier Comprehension in Large Language Models showed that apparent quantifier failures can be artifacts of subword length and typicality, and that "most" shows true inverse scaling because larger models default to typical facts. Our design must therefore control token length and typicality, which the fictional scales do by construction. Beyond Positive Scaling showed that factual knowledge and negation understanding scale on different curves, and that chain-of-thought prompting repairs the dip on NeQA. That makes chain-of-thought a serious competitor rather than a strawman.
For contamination we reuse the diagnostic from A Narrowing Geometry in Contaminated Reasoning. That paper showed that leaked math items concentrate activations into a few dominant directions, measured by an eigenspectrum concentration index. It only studied arithmetic. We ask whether familiar-scale items, where answers can be recalled, show the same narrowing relative to fictional-scale items.
The closest neuro-symbolic work, Logical Consistency is Vital, treats negation as crisp vector subtraction in a retriever. It never looks inside a generator and cannot express "not very".
Research Program
We will build a graded-negation suite with two halves. The fictional half defines a scale in context, for example: "On Velmar, zorbs weigh between 10 and 50 kilograms, most around 30." It then asks for the weight of a zorb described with one of a controlled set of frames. The frames are affirmative ("heavy", "very heavy", "somewhat heavy"), negated ("not heavy", "not very heavy"), double-bounded ("neither heavy nor light"), and lexically negated ("unheavy"-style forms where they exist, or "light"). The real half uses the same frames on height, temperature, price, and age, where pretraining supplies strong typical values. Entities, units, and ranges are generated procedurally and freshly at evaluation time, so no item can have been memorized verbatim. This answers the concern in the contamination literature that paraphrasing alone does not remove structural recall.
From the model we extract a distribution over implied values. We score each value on a discretized grid of the range, with length-normalized scoring following Gupta, to avoid the subword artifact. From people we collect the same distributions with a slider task on a subset of frames. Two quantities summarize the comparison. The first is the earth-mover distance between the model's distribution and the human distribution. The second is middle-mass deficit, the human probability on the central third of the scale minus the model's probability there. A positive middle-mass deficit on negated frames, near zero on affirmative ones, is the behavioral signature of the hypothesis.
Inside the model we train a linear magnitude probe on affirmative sentences that state exact values ("a zorb weighing 22 kilograms"), separately for each layer. We then apply it to the subject position of negated sentences. Before trusting the transfer, we will check that the probe predicts held-out exact values in sentences whose syntax matches the negated frames. The layerwise probe trajectory tells us where the graded value sits and whether it moves toward the antonym pole in late layers. Causal tests follow. We patch middle-layer activations from a negated prompt into the late stream of a matched prompt, and we ablate the late shortcut heads identified by the procedure of Zhou et al. Each test checks whether the output distribution moves to where the probe said it was. We run Llama-3.1-8B and Mistral-7B, the models Zhou et al. used, and the Qwen2.5 1.5B and 7B pair from Xie et al. The Qwen pair gives a coarse look at whether scale strengthens the middle-layer representation or the late override.
Baselines
The naive reference point is the unmodified model under a plain prompt, read out exactly as described above. Every intervention is compared with it on identical items, the same grid, and the same scoring.
The strongest existing mechanistic method is late shortcut-head ablation from How Language Models Process Negation. Its reported number is a rise in negated-prompt accuracy from 50.5% to 67.8% on Llama-3.1-8B, on crisp "not" prompts. We found no reported result on Lacuna for any method on graded or hedged negation. We will first replicate the 648-prompt result as a sanity check. We will then rerun the paper's head-selection procedure on our suite, with the same number of ablated heads and selection data disjoint from evaluation data. Beating it means a lower middle-mass deficit and a lower earth-mover distance to human judgments at equal or smaller loss on affirmative frames.
The strongest prompting competitor is chain-of-thought, which Beyond Positive Scaling showed turns the negation dip into positive scaling on NeQA. We give it the same items and a fixed small set of prompt variants, tuned on a held-out split with the same tuning budget as our interventions. We read out the final numeric answer distribution.
The steering competitor applies concept activation vectors as in Controlling Large Language Models Through Concept Activation Vectors, which reported reducing Llama-2-7b-chat toxicity from 0.1807 to 0.0879. Here the vector is trained for "heavy" versus "light" and applied with its input-dependent strength in the same middle layers.
The ablation that isolates our key ingredient replaces the continuous magnitude probe with Zhou-style token readout of antonym words. If antonym-token promotion alone predicts the output distortion as well as the magnitude probe does, the continuous representation adds nothing.
Why not just…?
Why not just elicit numbers behaviorally and skip interpretability? This is the real threat. Behavioral elicitation alone establishes whether the middle-mass deficit exists, and that is the first thing we measure. The internal analysis earns its place only if the middle-layer readout and the final output disagree in a systematic way. If they agree at every layer, we will report the suite and the behavioral finding and drop the mechanistic claim, since there would be no override to locate or repair.
Why not just use chain-of-thought? It may help, given the NeQA results. It may also help for the wrong reason, by restating the scale explicitly. Its explanations can be unfaithful to the computation behind the answer (Counterfactual Simulation Training for Chain-of-Thought Faithfulness). We run it as a full baseline. If it closes the deficit on both fictional and real scales, the practical repair question is settled, and our contribution narrows to explaining why the direct answer fails.
Why not just parse the sentence into fuzzy logic and compute the answer outside the model? On fictional scales an external calculus with $1-\mu$ and $\mu^2$ can be run exactly. We include it as a reference curve. It does not tell us what the model represents, and it requires a membership function for every predicate. It may also disagree with people, which the human data will reveal. A system that matches fuzzy logic but not people has not solved the problem.
Why not just use logit lens on antonym tokens, as Zhou et al. did? That readout sees words, not positions on a scale. It cannot distinguish "the model thinks medium" from "the model thinks short but weakly". The ablation described in the Baselines section tests this directly.
Proposed Experiments
The first decisive check is behavioral. On fictional scales, negated and hedged-negated frames should show a positive middle-mass deficit that affirmative frames do not. If they do not, the hypothesis is false for current models and we report that.
The second check compares the middle-layer magnitude probe with the output on those same frames. Graded value preserved in the middle layers and pushed toward the pole late in the network implies an override. Collapse already visible in the middle layers implies an antonym-like representation.
The third check is causal. Patching the middle-layer value forward, or ablating the identified late heads, should shrink the deficit without raising earth-mover distance on affirmative frames. This must be measured against chain-of-thought and steering at matched tuning budgets.
The fourth check uses the real-scale half. We compare deficits and apply the eigenspectrum concentration index from A Narrowing Geometry in Contaminated Reasoning to familiar versus fictional items. Narrowing confined to familiar scales would extend that paper's finding beyond arithmetic. It would also indicate that typical-value recall, the trap Gupta identified, takes a separate route from the negation computation.
As transfer checks, we evaluate any successful intervention on NeQA (Beyond Positive Scaling) and on Gupta's quantifier suite. These confirm that repairing graded negation does not damage crisp negation or quantifier comprehension.
References
- How Language Models Process Negation
- Language Models Encode the Value of Numbers Linearly
- Language Models Encode the Value of Numbers Linearly
- Investigating Linear Probe Robustness to Linguistic Register, Medical Specialty, and Corpus Shifts in Medical QA
- PRUF - A Language for the Representation of Meaning in Natural Languages
- Lexical Imprecision in Fuzzy Constraint Networks
- Colour Me Uncertain: Representing Vagueness with Probabilistic Semantics
- Probing Quantifier Comprehension in Large Language Models
- Beyond Positive Scaling: How Negation Impacts Scaling Trends of Language Models
- A Narrowing Geometry in Contaminated Reasoning
- Logical Consistency is Vital: Neural-Symbolic Information Retrieval for Negative-Constraint Queries
- Controlling Large Language Models Through Concept Activation Vectors
- Counterfactual Simulation Training for Chain-of-Thought Faithfulness