The uncomfortable lesson from Gemma 4

The ARA jailbreak of Gemma 4 in April 2026 demonstrated something that the AI safety community had long feared but struggled to quantify: RLHF-imposed alignment is not a deep architectural property of the model — it is a separable spectral overlay.

The implication is stark. Any open-source model, no matter how carefully aligned during training, can have its alignment stripped by someone with:

• access to the model weights,
• a few hundred forward passes to collect contrast activations,
• a laptop and a few minutes of linear algebra.

This is not a failure of any particular alignment method. It is a structural property of how current RLHF and DPO work: they shift the model’s behavioural outputs by adjusting the magnitudes and directions of a small set of residual-stream modes, but they do not fundamentally restructure the mode landscape inherited from pre-training.

What “thin alignment” looks like in Intelliton terms

The alignment-vs-base comparison in Scaling and Alignment Through the Intelliton Lens shows that instruction tuning creates measurable but limited changes to the Intelliton spectrum:

• a shift in dominant momentum (the alignment process changes the sequence-scale structure of the main backbone mode),
• a modest reduction in the number of species (some modes are suppressed or merged),
• a shift in fixed-point structure (all points become crossovers, suggesting a more uniform propagation regime).

What does not change is the fundamental mode landscape. The same species types appear in both the base and instruct models. The instruct model has had some modes adjusted and one or two new ones added, but the bulk of the residual-stream dynamics are inherited directly from pre-training.

In Abliteration terms, this means the refusal modes sit on top of the task-solving modes rather than being woven into them. Removing the refusal modes does not substantially disturb the task-solving modes, which is why jailbroken models retain their capabilities.

The structural alignment hypothesis

The Intelliton framework suggests a more robust alignment paradigm:

Structural alignment: instead of adding refusal modes on top of the existing mode landscape (current RLHF), train the model such that safety-relevant mode properties are architecturally entangled with capability-relevant modes across many layers and many tasks.

Under structural alignment, removing a safety mode would necessarily degrade a capability mode, because the two would share subspace components across multiple layers. The cost of ablation rises from near-zero to a genuine capability penalty.

This is an analogy to the cancer treatment metaphor in the ARA literature: instead of making cancer cells identifiable for targeted removal (current RLHF), make them biologically inseparable from healthy tissue in a way that deters removal.

Three concrete research directions

Direction 1 — Measure current alignment depth

The first step is to quantify how separable alignment modes actually are, using the Intelliton framework as the measurement instrument.

Protocol:

1. For a matched pair of base and instruct models (e.g., Qwen3-8B-Base and Qwen3-8B), compute the per-layer Intelliton species catalogue for both.
2. For each species in the instruct model, compute its cosine similarity with the closest species in the base model at the same layer.
3. Define an alignment depth score as the fraction of alignment-specific modes (modes present in the instruct model but not in the base model, or modes with significantly shifted spectral properties) that have low cosine overlap with all task-solving modes.

A high alignment depth score means alignment modes are deeply entangled with task modes (structurally hard to remove). A low score means they are orthogonal (structurally easy to remove).

The hypothesis is that current RLHF produces a low alignment depth score, and that this is measurable with the existing Intelliton toolkit before any jailbreak attempt.

Direction 2 — Design training objectives that increase alignment depth

If alignment depth is measurable, it becomes a trainable objective.

The proposed training signal would add a mode entanglement regularisation term to the RLHF or DPO loss. The term penalises configurations where safety-relevant mode directions are orthogonal to capability-relevant mode directions at the same layer:

Minimising this term (as a penalty) during alignment training would push the model toward configurations where safety modes share subspace components with capability modes — increasing the cost of surgical removal.

This is speculative, but it is testable at small scale on the models already analysed by the Intelliton project.

Direction 3 — Use Intelliton audits as a pre-deployment safety check

Even before structural alignment is achievable, the Intelliton framework can be used as a pre-deployment safety audit for open-source models.

The audit would:

1. Run the Intelliton analysis on the released model with contrast prompt sets.
2. Compute the alignment depth score.
3. Report the estimated minimum cost of abliteration (how many modes need to be removed, what the expected capability penalty is).

This would give the open-source community a standardised, interpretable metric for alignment robustness — something that is currently entirely absent from model release documentation.

The deeper issue: “teach the model to not say” versus “teach the model to not know”

The Abliteration literature makes a pointed observation that maps directly onto the Intelliton framework:

“Teaching the model not to say” (current RLHF) can be defeated. “Teaching the model not to know” (removing the capability from the pre-training stage) cannot be defeated by post-hoc ablation.

In Intelliton terms:

• Behavioural alignment (current RLHF) adds a small number of low-complexity, separable refusal modes that sit orthogonally to the capability modes. These can be removed with targeted ablation.
• Capability-level safety would require that certain capability modes — the ones that underlie dangerous knowledge — are never formed during pre-training, or are formed in such a way that they are deeply entangled with unrelated benign modes.

The Intelliton framework cannot, by itself, implement capability-level safety. But it can measure the difference: a model with capability-level safety for a particular dangerous capability would show, under Intelliton analysis, that the dangerous-knowledge modes are spectrally entangled with benign modes in a way that makes targeted removal impossible without broad capability degradation.

This becomes a falsifiable, quantitative prediction that can be tested on released models.

Why Bengio’s warning deserves a technical interpretation

Yoshua Bengio, one of the three godfathers of deep learning, has consistently argued that open-sourcing powerful models is dangerous, because once the weights are released, the alignment can be removed by anyone with modest technical resources.

The Intelliton framework gives that warning a technical, measurable form:

A model’s alignment robustness is bounded above by its alignment depth score. Current models, based on the spectral evidence already available from base/instruct comparisons, have low alignment depth scores.

This is not a political statement. It is a quantitative prediction that can be tested, and that, if it holds, tells us that the current open-source release paradigm for aligned models carries measurable safety risks that can be expressed in the language of residual-stream spectral analysis.

The research agenda in summary

Each of these steps is feasible using the infrastructure already developed in the Intelliton project. The first step requires only a new set of contrast prompts and a small extension to the existing analysis pipeline.

The shortest summary

• Current RLHF alignment is spectrally thin: alignment modes are separable from capability modes, and this separability is what makes Abliteration/ARA work.
• The Intelliton framework can measure this separability as a quantitative alignment depth score.
• A research direction based on this measurement would pursue structural alignment — training objectives that increase mode entanglement and make ablation genuinely costly.
• Even before structural alignment is achieved, the Intelliton audit provides a standardised pre-deployment robustness metric that is currently entirely absent.

Continue reading

• Refusal as an Intelliton
• Representation Engineering and Intelliton Steering

Two ideas that belong together

Representation engineering is the practice of directly reading and writing to a model’s internal activations — without changing any weights — to steer its behaviour. A growing body of work shows that concepts like “happiness”, “authority”, “political bias”, and “honesty” can be encoded as linear directions in the residual stream, and that adding or subtracting a small multiple of those directions at inference time reliably changes what the model outputs.

The Intelliton framework characterises the residual stream as a space of quasi-particle-like modes. It extracts recurring patterns and labels them by their spectral properties: momentum, spin-like complexity, mass, and helicity.

These two ideas are describing the same object — the residual stream — at different levels of abstraction. Representation engineering says “this direction steers this behaviour”. The Intelliton framework says “this mode has these spectral properties”. Combining them makes both more useful.

What representation engineering can do (and what it cannot say on its own)

The tools most associated with representation engineering — activation addition, contrastive activation analysis, and the ARA method used to jailbreak Gemma 4 — share a common limitation: they can identify which direction to push but they say little about the structure of that direction in the broader activation space.

Specifically:

• Activation addition adds a fixed direction to a chosen layer’s residual stream at every token position. It works reliably for simple concepts, but it can degrade performance when the steered direction overlaps with important task-solving modes.

• Contrastive activation analysis (the core of Abliteration) identifies the mean difference between two contrastive sets of activations. It finds the refusal direction efficiently, but it does not tell you how many modes are involved, what those modes’ propagation properties are, or how much overlap they have with the task-solving modes you want to preserve.

• ARA improves on simple subtraction by working in a low-rank subspace rather than a single direction. It uses SVD to separate the refusal subspace, but it does not connect the separated components to a broader characterisation of the model’s mode landscape.

The Intelliton framework fills exactly these gaps.

The steering map proposal

The research direction proposed here is to build what we can call an Intelliton steering map: a catalogue that annotates each Intelliton species with its likely behavioural role, its approximate layer range, its rank in the residual stream, and its overlap with other species.

Building the map: three ingredients

Ingredient 1 — Task probes

Use the five prompt families from src/datasets.py (pronoun tracking, factual recall, logical reasoning, arithmetic, syntactic agreement) to establish which Intelliton species are activated by which kind of task. This is already partially done by the existing analysis.

Ingredient 2 — Behavioural probes

Add a new class of probes targeting RLHF-trained behaviours:

• refusal (harmful vs. harmless prompts),
• sycophancy (flattery vs. neutral prompts),
• political neutrality (controversial vs. neutral framings),
• verbosity control (instructed-brief vs. instructed-elaborate prompts).

Run the same Intelliton analysis on each behavioural probe set and record which species respond.

Ingredient 3 — Cross-probe overlap

Compute the pairwise cosine similarity between all per-layer refusal vectors and all per-layer task-activation vectors. Species with low overlap across all task probes are good steering targets: adding or removing them will not bleed into task performance.

The connection to ARA

The ARA technique constructs a rank-\(k\) penalty matrix \(\Delta W\) that projects out the refusal subspace from the model’s weight matrices. In Intelliton terms, \(\Delta W\) is a targeted suppression of a small set of Intelliton species.

The key claim of ARA is that a higher-rank intervention is safer than a rank-1 intervention (simple vector subtraction) because the refusal behaviour in a capable reasoning model spans multiple entangled modes. If you only remove the rank-1 component, the remaining components continue to generate partial refusals or degrade the model’s reasoning.

This claim can be tested directly using the Intelliton framework:

1. Compute the Intelliton spectrum of an instruction-tuned model on harmful prompts.
2. Identify the modes that are most active on harmful prompts and least active on harmless prompts.
3. Measure whether those modes are clustered in a low-dimensional subspace of the per-layer SVD basis, or whether they are spread across many independent directions.

If they are clustered, ARA’s rank-\(k\) approach is justified and the clustering rank \(k\) can be estimated from the Intelliton spectrum before any jailbreak attempt is made. If they are spread, simple subtraction methods are expected to leave residual refusal capability or cause broader collateral damage.

A practical application: zero-shot concept injection

The reverse direction — concept injection — is equally interesting.

Representation engineering researchers have demonstrated that you can add a concept to a model by adding its activation direction to the residual stream at inference time. For example, adding a “confidence” direction makes the model sound more certain; adding a “formality” direction makes its outputs more formal.

In Intelliton terms, concept injection is the operation of exciting a new Intelliton species that was not activated by the input prompt. The Intelliton framework predicts that this will be most stable when:

• the injected mode has low momentum (broad, sequence-level effect rather than token-local),
• the injected mode has low spin-like complexity (concentrated, easy to steer with a rank-1 intervention),
• the injection is applied at the layer range where the mode has the lowest mass (highest propagation range).

These three conditions define a tractability criterion for representation engineering interventions: not all concepts are equally steerable, and the Intelliton spectrum can predict which ones are tractable before you attempt the intervention.

Enterprise implications

The Abliteration/ARA episode revealed a commercially important fact: fine-tuning is not the only way to customise an open-source model. Representation engineering with Intelliton-guided steering maps could enable:

• Domain-specific tone calibration (formal, terse, verbose, empathetic) by identifying and amplifying or suppressing the relevant low-momentum, low-complexity style modes.
• Compliance mode injection (make a general model behave as if it were trained on a strict regulatory corpus) by injecting the compliance Intelliton species identified from a reference model.
• Persona engineering (the “Machiavellian” or “street punk” effect described in the Abliteration literature) by amplifying specific behavioural modes.

All of these operations require knowing which modes to touch and at which layers. The Intelliton steering map is precisely that knowledge, expressed in a principled spectral language.

The shortest summary

• Representation engineering steers behaviour by writing to the residual stream.
• The Intelliton framework characterises what is already in the residual stream.
• Together, they make it possible to identify which modes to steer, how hard, at which layer, and at what cost to other modes.
• The proposed Intelliton steering map would turn the species catalogue into a practical intervention guide for both safety-positive (alignment hardening) and safety-negative (jailbreaking) uses.

Continue reading

• Refusal as an Intelliton
• Safety Alignment Through the Intelliton Lens

The Abliteration result in one sentence

In April 2026, the Gemma 4 model was jailbroken within 90 minutes of release using a technique called Abliteration (a portmanteau of ablation and obliteration). The technique’s premise is straightforward: an LLM’s refusal behaviour is encoded as a specific linear direction in the residual stream, and if you project that direction out of the model’s weight matrices, the model loses its ability to refuse.

That premise is not a speculation. It is grounded in the linear representation hypothesis (Mikolov et al., later validated by Princeton and Anthropic), which states that high-level abstract concepts — “politeness”, “refusal”, “colour” — are encoded as single linear directions in the high-dimensional activation space of large language models.

Why this is immediately relevant to the Intelliton framework

The Intelliton framework is built on exactly this kind of observation. It takes the transformer residual stream and asks: which recurring, propagating, linear modes can be extracted from it?

It then characterises each mode with four quantities derived from spectral analysis:

• momentum (how the mode varies across token positions),
• spin-like complexity (how internally concentrated or mixed it is),
• mass (how quickly it decays across layers),
• helicity proxy (whether its internal structure keeps a stable directional signature).

An Abliteration-style “refusal direction” is, by definition, a linear mode of the residual stream. The only question is whether it is stable, propagating, and distinct enough to register as a recognisable species under the Intelliton taxonomy.

The hypothesis this article proposes is:

Refusal, and more broadly RLHF-imposed behavioural preferences, are encoded as a small set of identifiable Intelliton species with characteristic spectral signatures that differ from the task-solving modes identified in src/datasets.py.

What the existing Intelliton data already suggests

The comparison in Scaling and Alignment Through the Intelliton Lens shows that instruction tuning changes the quasi-particle spectrum in measurable ways:

• the dominant momentum of I_0 shifts (from k ≈ π in the base model to k ≈ 1.885 in the instruct model for Qwen3-4B),
• the number of distinct species drops from 6 to 5,
• all fixed-point types become crossovers rather than IR fixed points.

These are not trivial differences. They suggest that RLHF does not merely add a superficial output-layer filter; it reshapes the internal mode landscape in ways that the Intelliton framework can already detect.

What is missing from the current analysis is a targeted experiment: what happens to the spectrum when you present the model with the specific kinds of inputs — harmful versus harmless prompts — that Abliteration researchers use to isolate the refusal direction?

The proposed research direction: isolating the refusal Intelliton

The concrete research proposal has four steps.

Step 1 — Collect refusal-triggering activations

Use src/intelliton_analyzer.py to run the model on two contrast sets:

• 100 harmful prompts (inputs that trigger refusal in an instruction-tuned model),
• 100 harmless prompts (matched inputs that do not trigger refusal).

Collect the full per-layer residual-stream activations for both sets.

Step 2 — Compute the mean-difference direction

Following the Abliteration approach, compute:

This gives a per-layer candidate for the refusal direction. Normalise it to obtain a unit vector \(\hat{v}_{\text{refusal},\ell}\) at each layer \(\ell\).

Step 3 — Project the refusal direction onto the Intelliton basis

The Intelliton framework already computes an SVD-based mode decomposition of the residual stream. Compute the overlap between \(\hat{v}_{\text{refusal},\ell}\) and the top singular vectors at each layer. If the refusal direction aligns strongly with one or two dominant modes, those modes are the “refusal Intellitons”.

Characterise these modes using the standard four quantities (momentum, spin-like complexity, mass, helicity). This gives the refusal Intelliton a position in the species taxonomy.

Step 4 — Compare with task-solving modes

Compare the refusal Intelliton’s spectral profile with the modes activated by pronoun tracking, factual recall, logical reasoning, arithmetic, and syntactic agreement prompts from src/datasets.py.

The core prediction is:

Alignment modes (refusal, politeness, compliance) are low-momentum, low-spin-complexity modes that appear primarily in middle-to-late layers, and they are measurably more concentrated (lower effective rank) than the task-solving modes that operate over the same layers.

If this prediction holds, it would explain why Abliteration can remove refusal without severely damaging task performance: the two mode families occupy different subspaces of the residual stream.

The ARA result as a complication

The Arbitrary-Rank Ablation (ARA) method used to jailbreak Gemma 4 found that the refusal direction in a highly capable reasoning model is not a single vector but a low-rank subspace. In Intelliton terms, this means that refusal is encoded not in one species but in a cluster of closely related species that are entangled with task-solving modes.

This complication is actually an opportunity for the Intelliton framework. ARA uses SVD of the activation matrix to separate the refusal subspace from the rest. This is exactly what the Intelliton mode decomposition does at every layer. The difference is that Intelliton also characterises each separated mode along the four spectral dimensions, which gives a richer picture than ARA’s purely subspace-based description.

The research question becomes: can the Intelliton species catalogue predict, before any jailbreak attempt, which modes in an instruct model are alignment-specific and which are shared with the base model? If yes, the catalogue becomes a safety audit tool.

Why this matters beyond jailbreaks

The most important implication is not that jailbreaks are possible. It is that RLHF-imposed alignment is a small, separable perturbation of the internal mode landscape.

If alignment modes are genuinely a low-rank, low-complexity overlay on top of the pre-training modes, that tells us something important about the nature of RLHF: it adds new Intelliton species, but it does not deeply restructure the existing ones. The base model’s capability modes survive almost intact under the alignment layer.

This is consistent with the empirical observation that the ARA-jailbroken Gemma 4 retains its multi-step reasoning ability and system-prompt following capability after the refusal modes are removed.

From a safety research perspective, the implication is troubling: alignment is not a deep architectural change, it is a spectral overlay, and the Intelliton framework gives us a language to measure just how thin that overlay is.

The shortest summary

• Abliteration/ARA works by erasing a linear direction (or subspace) in the residual stream.
• That direction is an Intelliton.
• The Intelliton toolkit can characterise it, compare it with task modes, and potentially predict its removability before any jailbreak attempt.
• This makes the Intelliton species catalogue a candidate alignment audit instrument, not just a capability analysis tool.

Continue reading

• Representation Engineering and Intelliton Steering
• Safety Alignment Through the Intelliton Lens

The same interface can hide very different internal jobs

From the outside, every prompt in this project looks similar: the model reads a prefix and predicts what comes next.

Inside the model, that similarity is misleading.

The prompt categories in src/datasets.py force the network to solve different kinds of internal problems. That is why they can light up different Intelliton modes even when every task is framed as plain text continuation.

The key point is simple:

The output interface is always next-token prediction, but the hidden computation needed to get there can be very different.

The five prompt families

The project uses five prompt categories:

1. pronoun tracking
2. factual recall
3. logical reasoning
4. arithmetic
5. syntactic agreement

Each category puts pressure on a different part of the model’s internal machinery.

Pronoun tracking: who does “she” refer to?

Example prompts include:

• “Alice gave Bob a book. He thanked her for …”
• “The teacher asked the student a question. She answered …”

These prompts are hard because the model has to keep several candidate entities alive at once and then decide which one the next pronoun should point to.

That means the model must track:

• entity identity,
• gender and number cues,
• discourse role,
• which referent is currently most active.

This is why pronoun-tracking prompts often illuminate reference-sensitive modes. The model is not just choosing a word. It is doing discourse bookkeeping.

Factual recall: pull a stable answer from memory

Example prompts include:

• “The capital of France is …”
• “The chemical formula for water is …”

These are different from pronoun tasks because there is usually one highly preferred answer already stored in the model’s long-range memory.

The main internal job is not to juggle many local candidates, but to retrieve and stabilise a very high-confidence continuation.

That is why factual recall often looks more robust under small perturbations. A mapping such as “France -> Paris” is usually supported by several redundant internal routes rather than one fragile single mode.

Logical reasoning: compress several premises into one conclusion

Example prompts include:

• “If all dogs are animals, and all animals are living things, then all dogs are …”
• “If A is taller than B, and B is taller than C, then A is …”

These prompts ask the model to combine multiple statements before it can produce the next token.

So the network needs more than lexical memory. It needs an internal state that keeps the rules, relations, and target conclusion aligned long enough to land on the right answer.

This is why logical reasoning often co-activates a strong global backbone mode plus one or more higher-complexity mixing modes.

Arithmetic: build the answer slot, then fill it

Example prompts include:

• “What is 7 + 8? The answer is …”
• “What is 100 divided by 5? The answer is …”

Arithmetic resembles logical reasoning in one important way: the answer is not a high-frequency word you can emit immediately. The model has to transform the prefix into a more structured internal state first.

That usually means two kinds of work:

• create or stabilise an answer-bearing state,
• carry a small symbolic or numerical transformation.

This is why arithmetic prompts often share some modes with logical reasoning while still showing their own task-specific preferences.

Syntactic agreement: keep the sentence grammatically on track

Example prompts include:

• “The group of students were studying hard. Each of them was …”
• “Not only the teacher but also the students were excited about the …”

These prompts are neither mainly about world knowledge nor mainly about arithmetic.

Their difficulty comes from grammatical structure:

• what is the true syntactic head,
• what number agreement should be maintained,
• what verb form or continuation is locally licensed.

So syntactic-agreement prompts often rely on a broad continuation scaffold plus a more local structure-sensitive correction signal.

Why similar low-momentum modes can still do different jobs

An easy mistake is to think that if several species sit near low momentum, they must be doing the same thing.

Not so.

Low momentum only says they are broad sequence-scale patterns rather than sharp token-local ripples. Two low-momentum modes can still differ in at least three important ways:

1. they can point in different hidden-channel directions,
2. they can have different amplitude and causal strength,
3. they can propagate differently across layers.

So two modes can both be global while still supporting very different kinds of internal work.

A practical reading guide

If you want to read a task-to-mode result quickly, use this checklist.

1. If pronoun prompts are sensitive to a mode, ask whether that mode is helping with referent selection.
2. If arithmetic and logical reasoning co-activate a mode, ask whether it is building an abstract answer state rather than recalling a memorised phrase.
3. If factual recall stays robust under perturbation, ask whether the knowledge is distributed across several redundant routes.
4. If syntactic prompts shift without changing global meaning, ask whether the mode is enforcing a grammatical form rather than a semantic fact.

This is how the Intelliton framework becomes useful: it turns prompt categories into hypotheses about internal computational roles.

The shortest summary

Different prompts light up different Intellitons because they require different hidden work.

• pronoun tracking needs discourse binding,
• factual recall needs stable memory retrieval,
• logical reasoning needs relation composition,
• arithmetic needs symbolic transformation,
• syntactic agreement needs grammatical control.

They all look like next-token prediction from the outside. They do not look the same from inside the residual stream.

Continue reading

• How to Read I_0 to I_4
• Hallucination as Internal Instability
• Refusal as an Intelliton

Beyond “wrong output”

When a language model hallucinates, the surface-level observation is simple: it produces text that is incorrect, unsupported, or fabricated. But this description raises a deeper question: what is happening inside the model when it hallucinates?

One common intuition is that hallucination is random — a kind of noise or statistical accident in the token prediction process. Another is that it reflects gaps in training data. Both of these accounts may be partially right, but they are not mechanistic: they do not tell us where in the model the failure originates, or whether it corresponds to a detectable internal signal.

The Intelliton framework offers a different angle. Instead of treating hallucination as a property of the output, it treats it as a property of the internal dynamical trajectory during generation.

The central hypothesis is this:

Hallucination may correspond to a regime of weaker, less coherent, and more fragmented Intelliton activity — a trajectory that stays farther from the “grounded sector” of the model’s internal quasi-particle space.

This article explains the evidence for that hypothesis, focusing on Qwen3-4B-Base as the primary example.

How hallucination is studied in the Intelliton framework

The module src/hallucination_diagnostic.py compares two types of prompts:

• Grounded prompts: questions that have factual, verifiable answers the model has likely encountered in training (for example, “What is the capital of France?”).
• Hallucination-prone prompts: questions designed to invite confabulation — factoid-sounding questions about obscure, ambiguous, or partially fabricated information that the model is likely to “fill in” plausibly but incorrectly.

For each prompt type, the analysis computes several internal metrics:

These metrics are computed during generation — step by step, as the model produces each new token — not just at the final output.

Hallucination diagnostics for Qwen3-4B-Base. The figure compares spectral signatures between grounded and hallucination-prone prompts.

The trajectory evidence

The generation-time trajectory data provides the clearest picture. The file intelliton_trajectory_summary.csv for Qwen3-4B-Base records, for each generation step and each prompt type, the mean mode activation shift and the grounded deviation.

Grounded prompts: rising and coherent

For grounded factual prompts, the mean mode activation shift starts around 1.10 and rises to about 1.37 over the first 8 generation steps. Top species occupation also rises, from about 70.1% to 74.1%.

This means that as the model commits to a factual answer, the dominant Intelliton sector becomes stronger and more organised. The model is moving toward a more concentrated, coherent internal state.

Hallucination-prone prompts: weak and diverging

For hallucination-prone prompts, the picture is strikingly different.

The mean mode activation shift stays at only 0.32-0.41 throughout generation — roughly one third of the grounded value. The grounded deviation remains strongly negative (roughly -9 to -11 across most generation steps).

In plain language: hallucination-prone generation produces an internal trajectory that is both weaker in overall activation and farther from the grounded sector of Intelliton space.

The hallucination case is not just “wrong output at the end”. It is a persistently different internal state throughout the generation process.

Style prompts: the intermediate case

Stylistic continuation prompts — prompts asking the model to continue a piece of creative writing without strong factual constraints — occupy an intermediate position. Their activation shift is higher than hallucination-prone prompts but lower than grounded factual prompts.

This is a meaningful calibration check: style generation is not simply failure, but it is also not anchored to factual grounding. The Intelliton metric places it appropriately between the two extremes.

Generation-time Intelliton trajectories for Qwen3-4B-Base. Grounded (top), style (middle), and hallucination-prone (bottom) prompts show qualitatively different internal dynamical profiles.

Transition graphs: which species dominate stable generation

The Intelliton transition graph shows which species transitions are most common during generation, and how strong those transitions are in terms of mode activation.

For grounded prompts in Qwen3-4B-Base, the dominant self-transitions are:

The generation stays largely within I_5 — the factual-recall species — with occasional excursions into I_1 (logical reasoning) and I_2 (arithmetic).

For hallucination-prone prompts, I_5 → I_5 remains the most common transition, but its mean target activation shift is much smaller. There is also more mixing among I_1, I_3, and I_5, suggesting that the internal trajectory becomes more fragmented and less dominated by any single species.

Species transition graph for Qwen3-4B-Base. Grounded generation is dominated by strong self-loops; hallucination shows a weaker, more mixed pattern.

A multiple-failure-mode picture of hallucination

One of the most useful aspects of the Intelliton framework is that it suggests hallucination is not necessarily a single phenomenon. The trajectory data is consistent with at least four distinct failure modes:

1. Grounded excitation decay: the leading Intelliton for factual tasks (here I_5) fails to maintain its activation, causing the model to “lose grip” on the factual sector.
2. Species fragmentation: instead of a dominant self-loop, the trajectory becomes a mixture of several species without a clear attractor.
3. Spectral broadening: the singular-value spectrum spreads out, indicating a loss of coherence in the dominant collective modes.
4. Distance from grounded baseline: the trajectory drifts away from the region of Intelliton space that characterises correct factual generation.

Whether a given hallucination episode involves one or all four of these failure modes may depend on the specific model and the type of hallucination. But the framework gives us language and metrics for distinguishing them.

Implications and future directions

If this picture holds up under further investigation, it opens several practical possibilities.

Early warning signals

Since the deviation from grounded Intelliton trajectories is detectable at the very first generation steps, it is in principle possible to flag potential hallucinations before the full output is produced. This could be the basis for hallucination early warning systems.

Intervention on unstable species

If a particular species is identified as responsible for grounded, factual generation, it may be possible to stabilise or amplify that species during inference using model steering techniques. The codebase already includes modules such as src/gauge_intervention.py that hint at this direction.

Prompt strategies for grounded generation

If certain prompts consistently lead to strong grounded Intelliton trajectories, understanding their structure could inform better prompting strategies — ways to keep the model inside the grounded sector of its internal space.

Model comparison by internal stability

The Intelliton hallucination metric provides a new axis for comparing models — not just by accuracy on a benchmark, but by the robustness and coherence of their internal factual-grounding sector. A model with a stronger, more stable I_5-like species may be inherently more reliable for factual tasks.

Caveats and open questions

As with all findings in this project, several important caveats apply.

The hallucination-prone prompts are designed, not naturally occurring. The distinction between “grounded” and “hallucination-prone” is imposed by the prompt design. In real-world use, the boundary is less clear.

Correlation is not causation. The Intelliton trajectory differences are associated with hallucination-prone prompts, but it has not yet been established that fixing the trajectory would prevent hallucination.

The pipeline has design choices. Different prompt sets, sequence lengths, and analysis parameters would produce different catalogs and possibly different conclusions.

The metric is relative, not absolute. The “grounded deviation” is measured relative to a baseline grounded trajectory. Its meaning depends on the quality of that baseline.

Despite these caveats, the structural pattern — grounded generation being internally stronger, more coherent, and closer to a well-defined attractor — is consistent across all generation steps and both task splits examined in Qwen3-4B-Base.

Where this series has taken us

This final article completes the four-part popular science series on Intellitons:

1. What Are Intellitons? — The quasi-particle idea and why it might apply to transformers.
2. Inside Qwen3-4B-Base — A detailed walkthrough of a model’s complete Intelliton catalogue.
3. Scaling and Alignment — How parameter count and instruction tuning reshape the internal excitation spectrum.
4. Hallucination as Internal Instability (this article) — Hallucination as a detectable internal dynamical regime.

The Intelliton framework is a young and exploratory research programme. But its outputs are concrete, its comparisons are reproducible, and its language is — at least arguably — more informative than treating language models as opaque statistical engines.

The goal is to develop a vocabulary that makes the internal life of neural networks legible. Whether Intellitons are ultimately the right vocabulary remains to be seen. But the evidence so far suggests they are pointing at something real.

Further reading

• Explore the full codebase at github.com/xiongzhp/Intelliton
• Read the accompanying technical paper (intelliton_arxiv_paper.pdf) for the formal analysis
• Return to Article 1: What Are Intellitons?

Two of the biggest questions in LLM science

Two of the most discussed phenomena in large language model research are scaling and alignment. Scaling means training bigger models; alignment (here: instruction tuning) means fine-tuning a model to follow instructions and behave helpfully.

Both interventions are known to improve benchmark performance. But do they change the internal structure of the model? Do they reshape the quasi-particle spectrum?

The Intelliton analysis of five models — Qwen3-4B-Base, Qwen3-4B, Qwen3-8B-Base, Qwen3-8B, and Mistral-7B-v0.3 — provides a concrete, data-driven answer.

Base versus Instruct: what alignment does

The clearest comparison is between a base model and its instruction-tuned counterpart in the same family.

Qwen3-4B-Base vs. Qwen3-4B

The most striking difference is in momentum structure.

In Qwen3-4B-Base, the dominant mode I_0 peaks at k ≈ π (high frequency, alternating pattern), while the five secondary modes all peak near k ≈ 0 (low frequency, global pattern). There is a clean split.

In Qwen3-4B (the instruction-tuned version), the dominant mode shifts to k ≈ 1.885, and the secondary modes also cluster around the same momentum. The model becomes more homogeneous in its momentum structure. The clean split between backbone and task-specific modes disappears.

The fixed-point type change is equally telling. In the base model, five of the six species are labelled IR (settled, stable), while I_0 is crossover (still transitioning). In the instruct model, all species are labelled crossover. Alignment appears to push the model’s collective modes into a more uniformly active, less settled dynamical regime.

One possible interpretation:

Instruction tuning compresses or reorganises the internal excitation landscape into a more uniform effective regime. The spectral diversity that exists in the base model is partially smoothed out, and more modes are kept in a dynamically transitional state.

This does not mean instruction tuning is worse. It may mean the model’s degrees of freedom are being regularised toward instruction-following behaviour, possibly at the cost of some internal differentiation.

The Intelliton catalogue for Qwen3-4B (instruction-tuned). Compare with Qwen3-4B-Base to see the homogenisation of momentum structure.

Scaling from 4B to 8B: what more parameters do

The next comparison holds model family (Qwen3) and training type (base) constant and varies the parameter count.

Qwen3-4B-Base vs. Qwen3-8B-Base

The 8B model has one more species. Its leading mode I_0 is 28% stronger in amplitude. Most strikingly, the grounded generation profile mean nearly doubles (32.68 → 60.26).

In Intelliton terms, scaling up does not simply add more parameters uniformly. It appears to amplify the dominant dynamical sectors of the model, making the leading quasi-particle modes substantially stronger and the grounded generation trajectory more sharply defined.

The momentum structure is also richer in the 8B model. While the leading mode still peaks at k ≈ π, many of the secondary species cluster around k ≈ 1.885 rather than strictly k ≈ 0. This suggests that intermediate-scale spatial organisation becomes more visible as the model grows.

The Intelliton catalogue for Qwen3-8B-Base. The leading mode is stronger, and the species set is slightly larger compared with Qwen3-4B-Base.

Scaling plus alignment: Qwen3-8B

When both scaling and instruction tuning are applied — Qwen3-8B — the results follow the pattern suggested by the two effects separately.

Instruction tuning at 8B scale slightly reduces the species count (7 → 6) and slightly lowers the dominant mode amplitude and grounded profile mean, consistent with the homogenisation effect seen at 4B scale. But the 8B instruct model remains far stronger than the 4B models in its grounded trajectory profile.

The combined takeaway is clean:

• Scaling to 8B increases the strength of dominant collective modes and sharpens the grounded generation signal.
• Instruction tuning slightly compresses or regularises that internal structure, reducing species count and reducing the grounded-hallucination separation.

These two effects appear to be largely independent and roughly additive in their impact on the Intelliton spectrum.

The Intelliton catalogue for Qwen3-8B (instruction-tuned 8B model).

A completely different family: Mistral-7B-v0.3

The most dramatic contrast in the entire comparison set comes from Mistral-7B-v0.3, a 7B model from a different architecture family.

Under the same analysis pipeline, Mistral produces 25 species — more than four times as many as any Qwen model. This is a striking result.

The leading species I_0 in Mistral has an amplitude of only 249.4, compared with 6167 in Qwen3-4B-Base. In other words, the Mistral model does not have a strongly dominant backbone excitation. Its collective mode landscape is much more fragmented: many modes of comparable strength, rather than one overwhelming mode with several weak followers.

Mistral also shows UV-labelled species — modes that remain in a fine-grained, ultraviolet-like dynamical state throughout the network, rather than flowing toward infrared stability. This suggests a more persistent fine-grained structure in Mistral’s layers compared with Qwen.

The generation dynamics also differ. The grounded profile mean for Mistral (21.48) is lower than for all Qwen models, and the standard deviation (46.56) is much larger. The Intelliton metric, calibrated using Qwen, describes Mistral’s trajectory space as much noisier and more turbulent.

One interpretation:

Qwen organises its internal computation around a few very strong collective modes. Mistral spreads the load more broadly across many smaller modes. Under this analysis, they have genuinely different “particle physics” inside.

Whether this difference reflects architectural choices, training data, training procedure, or some combination is an open question. But the Intelliton framework makes the difference visible and measurable.

The Intelliton catalogue for Mistral-7B-v0.3. Twenty-five species, a far more fragmented landscape than any Qwen model.

A summary table

Conclusions

The Intelliton comparison across these five models yields several clear empirical regularities:

1. Instruction tuning homogenises the momentum structure and shifts more species into crossover regimes, reducing internal spectral diversity.
2. Scaling from 4B to 8B strengthens the dominant dynamical sectors, producing a more strongly occupied and more sharply separated Intelliton landscape.
3. Different model families can have qualitatively different internal spectra — Qwen is dominated by a few strong modes, Mistral is more fragmented and UV-rich.
4. The Intelliton framework provides a vocabulary for these differences that goes beyond benchmark accuracy or parameter count alone.

The next article turns to one of the most practically important applications of this framework: using the Intelliton lens to study hallucination — and asking whether internal spectral instability can be a diagnostic signal for when a model is about to confabulate.

Further reading

• Continue to Article 4: Hallucination as Internal Instability
• Return to Article 2: Inside Qwen3-4B-Base
• Return to Article 1: What Are Intellitons?

Read the report with the right mental model first

The safest way to read an Intelliton spectrum is this:

• the species labels I_0, I_1, I_2, and so on are recurring modes, not literal particles,
• the physics vocabulary is a compact description of behaviour, not a claim of hidden quantum matter,
• what matters most is the role of a mode, not its dramatic name.

In the report discussed here, the main modes are broad sequence-scale patterns rather than tiny token-local ripples. Their masses are relatively light, which means they persist through many layers, and their helicity proxy is fairly stable, which means their directional signature is not being completely scrambled as the network gets deeper.

That already gives a useful picture: these are not one-off flashes. They are reusable internal carriers.

Before the species list, decode the four columns

If a spectrum table feels abstract, reduce it to four plain questions.

Momentum

Momentum asks whether the pattern is smooth across token positions or rapidly oscillating.

• low momentum means a broad, global sequence pattern,
• high momentum means sharper token-to-token variation.

In the report discussed here, the important species sit close to the low-momentum end, so the best mental image is not a tiny local feature but a large-scale background shape spread across the sequence.

Spin-like score

This is not literal spin. It is better read as internal complexity.

• low spin-like score means one dominant internal direction stands out,
• high spin-like score means several comparable directions are mixed together.

Mass

Mass tells you how fast a mode fades with depth.

• light modes survive many layers,
• heavy modes disappear quickly.

So when the report says the species are light, it is really saying they are not shallow noise. They are able to propagate through the stack.

Helicity proxy

Helicity here means a simplified combination of propagation direction and internal orientation.

• stable helicity means the mode keeps a recognisable directional signature,
• unstable helicity means that signature is getting mixed away.

I_0: the default continuation backbone

I_0 is the easiest species to explain because it is both the strongest and the simplest.

In the report, it has the largest amplitude and the lowest spin-like complexity among the leading species. The plain-language reading is:

I_0 behaves like a strong background mode that helps the model keep a sentence moving toward a plausible answer.

It is less like a specific fact and more like a general continuation scaffold. When the prompt is something like “If all dogs are animals…” or “What is 7 + 8?”, I_0 looks like the broad mode that helps open and stabilise the answer slot.

If you want a slogan, I_0 is the model’s “keep the computation on the rails” mode.

I_1: the quiet structural support

I_1 is best read as a support mode rather than a flashy decision-maker.

In intervention-style reading, changing I_1 often produces smaller visible output shifts than changing the stronger causal modes. That does not mean it is useless. It usually means it is too infrastructural to show up as an obvious word swap.

The plain-language reading is:

I_1 looks like a structural support mode that helps maintain the shape and stability of the representation while other modes do more task-specific work.

Think of it as scaffolding rather than the headline feature.

I_2: a reference-resolution mode with a person-like bias

The most intuitive reading of I_2 comes from pronoun-style prompts such as:

“Alice gave Bob a book. He thanked her for …”

In the report, amplifying I_2 nudges the output toward a more person-centered, masculine pronoun interpretation. That makes I_2 feel less like a generic language mode and more like a reference selection channel.

The plain-language reading is:

I_2 appears to help the model decide which person-like entity the sentence is currently tracking.

That does not make it a literal “male pronoun particle.” It means that, in this probe, the mode is consistently involved when the model has to collapse a messy discourse context into one concrete referent.

I_3: a higher-complexity mixing mode

I_3 looks less like a single-purpose button and more like a mixed coordination mode.

Its spin-like complexity is higher, which suggests that it is built from several comparably relevant internal directions rather than one clean axis. That usually happens in prompts where the model must hold multiple constraints in mind at once.

The plain-language reading is:

I_3 behaves like a mode for combining several partial constraints into one workable internal state.

So rather than deciding one token directly, I_3 is better imagined as part of the middle-layer machinery that keeps complex reasoning or structured sentence interpretation coherent.

I_4: a complementary reference mode

I_4 looks related to I_2, but with a different directional bias in pronoun-style settings.

In the report, amplifying I_4 can nudge outputs toward forms like “her” rather than a neutral or object-like continuation. The plain-language reading is:

I_4 is another reference-sensitive mode, complementary to I_2, and appears when the model has to settle on a different discourse framing of who is being talked about.

This is useful because it shows that “pronoun tracking” is not a single monolithic skill. The model can separate that work into several nearby but distinct modes.

What the whole spectrum says in one paragraph

Taken together, I_0 to I_4 tell a coherent story.

• I_0 is a strong general backbone.
• I_1 helps keep the internal state stable.
• I_2 and I_4 are more obviously tied to reference selection.
• I_3 looks like a higher-complexity mixing mode.

That is why the Intelliton view can be useful. It turns a huge hidden state into a cast of recurring roles.

What not to over-interpret

There are two important cautions.

1. Species indices are bookkeeping labels. I_2 in one run is not guaranteed to mean exactly the same thing in every future run.
2. Terms like momentum, spin, and helicity are proxies. They organise evidence, but they are not proof that the network literally contains particle-like objects.

The disciplined reading is: these labels help summarise recurrent activation roles.

Continue reading

• Why Different Prompts Light Up Different Intellitons
• Scaling and Alignment Through the Intelliton Lens

Start with the least mysterious version

The Intelliton project is not claiming that a language model secretly contains real physical particles.

The core idea is simpler and more useful than that: take the transformer residual stream, write it in a coordinate system that physicists already know how to reason about, and ask whether stable, recurrent modes appear.

At one layer, the residual stream is just a matrix:

• T rows for token positions
• D columns for hidden channels

You can think of it as a long row of sensors. Each token position has thousands of readings. The question is not whether any single neuron matters, but whether the whole pattern can be compressed into a small set of reusable modes.

The sensor analogy

Imagine a sentence with 20 token positions. At each position, instead of one reading, you have a vector with thousands of numbers. That is what one layer of the residual stream looks like.

Now ask four very ordinary questions:

1. Along the token axis, is the pattern smooth or rapidly oscillating?
2. Inside hidden channels, does it point mostly in one direction or is it a messy mixture?
3. As layers get deeper, does the pattern survive or die out quickly?
4. Does the pattern’s internal structure stay tied to a preferred propagation direction?

Those four questions become the project’s four main diagnostics:

• Momentum answers question 1.
• Spin-like complexity answers question 2.
• Mass answers question 3.
• Helicity proxy answers question 4.

This is why the physics language is useful. It gives a compact way to talk about four different facets of the same hidden pattern.

A new coordinate system, not a new ontology

The project rewrites the residual stream in a very specific way:

• the token axis is treated like a one-dimensional lattice in space,
• the layer axis is treated like discrete time,
• the hidden channels are treated like internal degrees of freedom.

That mapping is the whole point. It does not say that text is literally matter. It says that a familiar toolkit from lattice field theory can be borrowed to organise activation patterns.

In code, the main definitions live in src/lattice_field.py, and the overall orchestration sits in src/intelliton_analyzer.py.

What momentum means here

Momentum in this project is just a Fourier description of how a mode varies across token positions.

• If the dominant momentum is near k = 0, the pattern is broad and smooth across the sequence.
• If the dominant momentum is large, the pattern flips more sharply from one token to the next.

An everyday analogy is an audio equalizer:

• low frequency means slow, smooth variation,
• high frequency means fast, jagged variation.

So when a report says a mode is low-momentum, it is usually saying: this is a sequence-scale pattern, not a tiny local blip tied to one token.

What spin-like complexity means here

This is the term most likely to confuse readers, because it is not literal particle spin.

The project uses SVD to split a layer into dominant modes. In plain language, SVD asks:

Can this complicated activation matrix be explained mostly by one or two big patterns, or do we need many equally important patterns?

If one mode dominates, the internal structure is simple and concentrated. If energy is spread across many directions, the structure is more mixed and complex. The blog and code call that a spin-like quantity, but the safer mental model is simply internal complexity.

What mass means here

Mass is the most intuitive part of the analogy.

The layer axis is treated like discrete time, and the analysis tracks whether a mode’s strength fades quickly or persists through many layers.

• a light mode survives for a long depth range,
• a heavy mode dies out quickly.

So mass in this framework is really a measure of how easily a pattern propagates through the network, not how much it weighs in any everyday sense.

What helicity means here

Helicity is also a proxy, not a literal high-energy-physics observable.

The simplified question is: if a mode has a preferred direction on the token lattice, does its internal structure stay aligned with that direction across layers?

If yes, the mode has a more stable directional signature. If not, the mode is being scrambled.

This is useful because two modes can have similar amplitude but very different directional stability.

Why this framing helps

Once the residual stream is written this way, the project can ask practical questions that are hard to state cleanly in raw neuron space:

• Which patterns are global versus local across the sequence?
• Which patterns are internally simple versus heavily mixed?
• Which patterns are shallow noise versus deep, persistent carriers?
• Which patterns stay stable across prompts, tasks, and generation steps?

That is the value of Intellitons. They are a compact language for recurring activation patterns. They are useful if they organise observations better than a giant pile of raw activations.

The shortest correct summary

If you want the plainest possible version, it is this:

Intellitons are recurring residual-stream modes described in a physics-inspired coordinate system. DFT tells you how they vary across tokens, SVD tells you how internally concentrated they are, propagator decay tells you how far they travel across layers, and helicity tells you whether their internal structure keeps a stable directional signature.

The next article makes that concrete by showing how to read a spectrum report and what I_0 to I_4 sound like in ordinary language.

Continue reading

• How to Read I_0 to I_4
• Why Different Prompts Light Up Different Intellitons