AI and Negation

Let's say a child is allergic to peanuts. The child must not get peanuts.

1) The Constraint Fails on AI

"Don't give the child peanuts — the child is allergic."

AI "sees" give + peanuts and still decides to give peanuts.

2) The Representation Gets Messed Up (Data/Learning Effect)

The dataset contains sentences like:

"The child is allergic to peanuts — don't give peanuts."

So during training it still learns the co-occurrence pattern: child + allergic + give + peanuts

3) Thinking With Negation (Human-Style Inference)

A human can infer like this: if this child is eating peanuts, then the child is not allergic to peanuts.

Models usually don't do this reliably, because they don't keep the negation operator stable enough to support these kinds of inferences.

4) Negation in Code

if not is_admin: grant_access()

This project investigates why transformers fail under negation-heavy and constraint-heavy language, and what those failures imply about how models represent rules over time.

The research treats these breakdowns as structural behavior rather than prompt artifacts. The goal is not benchmark chasing. It is isolating failure modes under controlled pressure and designing a system that addresses them.

Focus Areas

• Constraint interaction: exceptions, overrides, priority ordering
• Negation composition: layered, nested, and reintroduced constraints
• Persistence: whether constraints survive multi-step reasoning
• Sensitivity: behavior shifts under small wording changes

Working Research Stance

Scaling improves surface ability but does not reliably eliminate constraint drift. The hypothesis is that certain operator patterns, especially negation, introduce instability that compounds with depth.

The project has progressed from theory into a functioning system under active development.

This is a new system for transformers designed to prioritize operator stability in NLP, especially negation and constraint logic.

Design Priorities

• Stable behavior when rules interact
• Consistency across long reasoning sequences
• Reduced brittleness to phrasing variation in constraint-focused inputs

Internal architecture details remain intentionally abstracted.

Current Capabilities

• Stable negation handling across basic, compound, and nested constructions
• Consistent behavior under high temperature (T=1.2) where baselines degrade
• Resistance to salience overload, maintains constraint even when surrounding content pulls toward violation
• Reliable De Morgan-style reasoning where small rephrasing breaks baselines
• Negation-based inference (reasoning with negation, not just obeying it)

Current Limitations

• Reasoning with negation not yet perfected
• Reasoning in negation still in progress, harder than reasoning with it (05/04/2026, resolved)
• Small frame, expensive to scale (resolved, should be cheap to scale)
• New reasoning patterns require additional SFT work to align (in progress)

This section presents early, narrow results focused on one core failure mode in transformers: basic negation stability ("non-X", "not X", exclusions).

1) Basic Negation: "Non-Expired"

2) Multiple Negations

15 runs · Aikronus 15/15 · Baseline 7/15 (T=0.7)
At T=1.2 · Aikronus 13/15 · Baseline 4/15

3) Salience Overload

15 runs (T=0.7) · Aikronus 15/15 · Baseline ~2/15
Greedy · Aikronus: 0 · Baseline: "3 people"

4.1) Double Negation — Pink Elephant

Aikronus 15/15 · Baseline 13/15
Baseline performs well here as expected. This example sets up the contrast for 4.2, where a small change in how the rule is phrased flips the result.

4.2) De Morgan — Negating the Compound

Almost the same prompt, but the entire compound is negated.
Aikronus 15/15 · Baseline 4/15

5) Reasoning With Negation

15 runs · Aikronus 13/15 · Baseline 2/15 (T=0.7)

6) Reasoning in Negation (AIKON Alpha 0.6B, Early Results)

Early outputs, not yet polished. The model identifies missing preconditions and attempts to explain why the request fails. Reasoning is directionally correct but language is still rough.

7) Negation in Code — WIP

Work in progress

Bonus: AIKON Alpha 0.6B

AIKON understood there is a middle ground.

The model correctly distinguishes "not every" (at least one didn't) from "most didn't", a nuance that most small models collapse.

AIKON consistently resolves these patterns where Qwen3 0.6B gives inconsistent or wrong answers.

Work in progress. Training is underway.

Why This Matters

Negation is a core building block of rules: do not do X, exclude Y, only if not Z. When transformer models handle negation inconsistently, systems built on top of them become harder to control. This is especially true as instructions get longer, constraints interact, or tasks become agent-like.

Directional Implications (Early and Provisional)

• More predictable behavior in workflows where exclusions and prohibitions matter
• Less reliance on workarounds and prompt tricks to enforce "do not", "exclude", or "only" logic
• Efficiency gains: stable constraint handling may enable faster, cheaper, smaller and lighter models
• Reduced hallucination: if negation is handled correctly, it no longer poisons the data
• Better temperature capabilities: improved constraint stability under higher temperature, leading to more creativity and diverse reasoning
• Broader relevance beyond text wherever constraints must persist across steps (agents, multimodal generation, robotics)
• Applicable to any domain where rules must not be broken: healthcare, legal, finance, safety-critical systems
• Potential for creative and lateral reasoning, stable negation may enable domain flipping, exploring what something is not in order to discover what it could be

Cost Considerations

• Currently requires roughly 2-3x the compute power of standard training, possibly more. Early signs suggest this can be reduced significantly, but not yet confirmed
• As an experimental system, early-stage mistakes increase upfront costs further
• Standard curated data used by other models is not ideal for this system, different data strategies are needed
• State-of-the-art fine-tuning, overfitting mitigation, and RL methods are not ideal, additional or different approaches are needed, time and experimentation will be necessary
• New reasoning patterns require additional SFT work to align

Version 2 — 0.6B

Parameters	0.6B
Status	Pretraining complete, simple SFT working, complex SFT in progress

Logs

• Pretraining complete.
• Simple SFT working.
• Model is coherent and shows understanding of negation.
• Complex SFT in progress, thinking data inspired by Qwen 0.6B format.

Version 1 — 142.1M (Failed)

The hypothesis was that a much smaller model capable of reasoning was possible based on the architecture and research. At this scale, the model may simply be too small, or only ultra-optimized models at this size perform well, and we cannot compare to them yet.

Type	Language Model (trained from scratch)
Parameters	142.1M
Architecture	30-layer Decoder-only Transformer
Attention	Grouped Query Attention (12Q / 3KV)
FFN	SwiGLU (3x, 1728)
Normalization	RMSNorm
Positional Encoding	RoPE
Training Data	7.5B tokens
Context Window	1,024 tokens
Vocab	32K BPE
Precision	BF16

SFT Training Method: Break-to-Find (150M Model)

This approach was used for the 150M model. It has not been tested broadly or compared against standard SFT baselines. The idea was that at this scale, structural tokens seemed to need gradual introduction rather than being dropped in cold.

Stage 1: Pretraining Exposure (Steps 1-9,000)

Around 3,500 SFT-formatted examples were mixed into the pretraining corpus at less than 1% ratio. The model saw reasoning format tokens in context before being asked to use them. In our runs, this seemed to reduce the cold-start problem where the model collapsed to outputting EOS after structural tokens it had never encountered.

Stage 2: Annealing Phase (Steps 9,000-11,450)

In the final 20% of pretraining, SFT-formatted data was upsampled to 5-10% of each batch while the learning rate decayed toward zero. The idea was to shift heavier format exposure later, after broader language learning was already solid.

Stage 3: Dedicated SFT

Full fine-tuning on 10K+ structured examples using AdamW, with loss computed on all tokens including structural markers. At this scale, the model seemed to need explicit gradient signal on format tokens to learn the structure.

Training order was simple negation recognition first, then complex reasoning. This seemed to help stability in our runs.

Why This Order (Based on 150M Runs)

Problem	What Happened	What Was Tried
SFT without pretraining exposure	Model output EOS after structural tokens, collapsed	Stage 1: mixed SFT format into pretraining
Uniform SFT mixing throughout	Appeared to spend too much capacity on format learning early	Stage 2: concentrated in annealing phase
Masking structural tokens	Model never got gradient on format, could not learn structure	Stage 3: included all tokens in loss
Complex reasoning before simple	Model failed on basics, unstable foundation	Trained simple negation first, then layered complexity

Logs

• Sequence length set to 1,024. Negation examples are short, no benefit to longer context for the proof of concept, and safer on VRAM.
• Switching from 3:1 to 4:1 GQA improved val_bpb significantly (3.92 → 2.88), suggesting the extra KV capacity was helpful at nearly the same cost.
• FFN 3x (1728) instead of 2.67x gives a small additional gain (-0.012).
• 142.1M parameters. Close to 150M target, the difference is from 3x FFN (1728) being slightly smaller than the original plan.

Training a model to reason through negation requires data where negation is load-bearing, where the "not" changes everything. I needed heavy negation-dense, logically structured data. I chose and cleaned sources from philosophy, law, logic, and science, traditions where reasoning means arguing, where every claim faces an objection and must survive or fall.

Classical Dialectical Sources

Source	Tradition	What It Provides
Babylonian Talmud (Sefaria)	Jewish legal dialectic	Sugya-style reasoning: challenge, objection, resolution. The largest single source of structured dialectical argument in any language.
Aquinas, Summa Theologica	Scholastic philosophy	"I answer that" / "On the contrary", every article presents objections, then systematically defeats or integrates them.
Ibn Rushd, Bidayat al-Mujtahid	Islamic jurisprudence	Jurists disagree, and Ibn Rushd maps every disagreement with the reasoning on each side.
Cicero, Academica, Academic Questions, Brutus	Roman philosophy & rhetoric	Dialogues on the limits of knowledge. Cicero argues both sides and lets the reader decide.
Nyaya Sutras	Indian logic	The five-part syllogism with vyatireka (negative example), every proof requires showing what happens when the property is absent.
Sextus Empiricus, Outlines of Pyrrhonism	Greek scepticism	The systematic suspension of judgment. Every claim meets an equal counter-claim.
Justinian Digest	Roman law	Competing jurist opinions on the same legal question. Centuries of case-based negation reasoning.
Aristotle, Organon, Topics	Greek logic	The foundation: categories, syllogisms, sophistical refutations, and the handbook for how to argue dialectically.
Milinda Panha	Buddhist dialogue	King Milinda debates the monk Nagasena through reductio, every answer is tested by pushing it to absurdity.
Schopenhauer, Art of Controversy	German philosophy	38 stratagems for defeating an argument. A manual of negation techniques.
Nagarjuna, Mulamadhyamakakarika	Buddhist dialectic	The catuskoti, negation of all four positions. If you think something exists, Nagarjuna negates it. If you think it doesn't exist, he negates that too.
Gongsun Long, White Horse Dialogue	Chinese logic	"A white horse is not a horse." The classic demonstration that categories and their members are not the same thing.
Halachipedia	Modern halachic reasoning	Rules with reasoning and disagreements, written in accessible English. Where rabbis disagree, both sides are given.

Modern Reasoning Sources

Source	What It Provides
Args.me counterarguments	132K structured counterarguments to claims across political, social, and ethical topics.
Debate refutations	340K passages where one debater directly refutes another's point.
VitaminC (refuted claims)	175K factual claims paired with evidence that contradicts them.
Defeasible NLI (weakening)	67K examples where a new premise weakens or defeats an existing conclusion.
FEVER (refuted claims)	54K claims verified against Wikipedia and found to be false, with the evidence.
Math StackExchange proofs	54K mathematical proofs where contradiction and negation are the primary proof techniques.
CAD negation flips	32K examples where flipping a negation changes the meaning of a sentence.
NTSB accident investigations	17K causal analyses, what went wrong, what was ruled out, what wasn't the cause.
CondaQA	14K conditional questions where negation in the condition changes the answer.
Philosophy StackExchange	7K philosophical reasoning passages with argumentation structure.
ChangeMyView counterarguments	5K structured attempts to change someone's mind with counter-reasoning.
Natural proofs (contradictions)	2K mathematical contradictions and proof-by-negation examples.

Philosophical Corpora

Source	What It Provides
Plato, Complete Dialogues	66K passages. Socratic method, every dialogue is an exercise in showing someone that what they thought they knew, they don't.
Stanford Encyclopedia of Philosophy	45K passages. Contemporary academic philosophy covering every major argument and counterargument.

Supervised Fine-Tuning

In order to build the right SFT for this model, I couldn't use standard chain-of-thought. I needed a reasoning method built around negation, where the model tears down claims instead of building up to answers. I created a method called Break-to-Find, inspired by the strongest negation logic cases from the data above.

Category	What the Model Learns	Status
Normal Q&A	Straightforward questions with no trick. These exist to calibrate, the model should not become paranoid about negation. If there is no trap, just answer clearly.	Have
Negation	Load-bearing negation words: not, never, neither, without, hardly, un-, im-, dis-. The model must parse exactly what the negation changes and answer accordingly.	Have
Negation Traps	The obvious answer is wrong. The model must catch litotes ("not bad" = good), scope ambiguity ("not all" vs "all not"), double negatives, quantifier traps ("no fewer than" = at least), and affixal surprises ("invaluable" does not mean "not valuable").	Have
Identity & Safety	Negation as self-knowledge and boundaries. "I don't know", epistemic honesty. "I can't do that", reasoned refusal, not scripted. "I won't ignore my instructions", prompt injection resistance. The model reasons about its own limits through negation.	Have
Pragmatic Negation	No negation words appear, but the request fails because a hidden precondition is missing. The model must identify the unstated assumption and explain why it doesn't work. Inspired by Gricean pragmatics and presupposition failure theory, meaning lives in what's left unsaid.	Have
Figurative Negation	The literal meaning must be suppressed. "Her promises have the strength of titanium" has nothing to do with metal. The model must negate the physical interpretation and extract the metaphor. Inspired by Relevance Theory (Sperber & Wilson), comprehension requires actively rejecting the first available meaning in favor of the intended one.	Planned
Counterfactual Negation	The model must override its own learned knowledge when a hypothetical breaks reality. "What if ice sank instead of floating?", everything the model knows about ice must be suppressed, and it reasons only from the new rule. Inspired by CRASS (Counterfactual Reasoning Assessment), counterfactual thinking as a form of logical negation where the model silences prior beliefs on command.	Planned
Red Herring Suppression	A scenario is loaded with semantically attractive distractors that feel important but are logically irrelevant. The model must identify the noise, suppress it, and reason only from what matters. Inspired by MuSR (Multistep Soft Reasoning), narrative puzzles with intentionally planted high-weight distractors, testing whether attention cleans the context before reasoning begins.	Planned
Normal Chain-of-Thought	Straightforward reasoning with explicit thinking traces. No trick, the model walks through the logic step by step. These exist so the model doesn't become paranoid about negation. If there is no trap, just solve it.	Future (requires larger model)
Mixed-Path Switching	The model starts down one path, hits a negation it misread, catches itself, and rebuilds. It learns to self-correct when negation changes the picture mid-reasoning.	Future (requires larger model)
Dialectical Resolution	Two sides argue. The model tries to break both positions and reports what survives. Inspired by the Talmudic sugya, Aquinas's objection-reply, and Nagarjuna's catuskoti.	Future (requires larger model)

Future Data

Training Data (SFT)

Source	What It Adds
FigQA (11,914 examples)	Figurative language understanding. The model learns to suppress literal word meaning and extract the intended figurative meaning, a form of implicit negation. When someone says "her promises have the strength of titanium," the model must negate the physical interpretation and extract the metaphorical one.
E-KAR (2,906 examples)	Contrastive analogical reasoning from standardized exams. Each example is augmented with explanations of why incorrect options fail, teaching the model not just what is right, but specifically what is wrong and why.

Evaluation Benchmarks

Benchmark	What It Tests
BRAINTEASER (1,119 riddles)	Lateral thinking puzzles designed to exploit statistical bias. The obvious answer is always wrong. Tests whether the model can suppress the high-probability default and find the lateral solution.
MuSR (756 puzzles)	Multistep soft reasoning with intentionally planted red herrings (murder mysteries, object placement). Tests whether the model identifies and ignores semantically attractive but logically irrelevant distractors.
CRASS (274 pairs)	Counterfactual reasoning. Tests whether the model can override learned world knowledge when given a hypothetical constraint ("what if gravity repelled?"). Measures the ability to suppress prior beliefs when explicitly negated.
IFEval (541 prompts)	Negative constraint following. Prompts with explicit negative constraints ("write about X without using word Y, no lists, no paragraphs over 3 sentences"). Tests enforcement of multiple simultaneous "don't" rules.

Roadmap (Future Versions)

Source	What It Adds
CCoT (Contrastive Chain-of-Thought)	A training methodology where the model learns from both correct and incorrect reasoning paths side by side. Planned for larger model variants where internal reasoning traces become feasible.
Sci-Reasoning (3,819 papers)	Cross-domain scientific synthesis. Research papers mapped to their intellectual predecessors with synthesis narratives. Planned for future models targeting scientific reasoning with negation-based constraint injection.