🤖 AI

Relational Intervention During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and a Structure x Register Factorial

This study demonstrates that during functional collapse in a small language model, a relational intervention combining specific structural elements (acknowledgment, absolution, agency restoration, unconditional acceptance) with a first-person register uniquely restores persistent behavior, revealing a dissociation where attention is driven by lexical surprise, emotional state by relational structure, and actual behavior by the interaction of both structure and register.

Original authors: Franco Santana, Horacio Vico

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Franco Santana, Horacio Vico

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot trying to fix a broken machine. Every time it tries, the machine breaks again. The robot keeps trying, getting more confused and stuck in a loop, like a hamster running on a wheel that's falling apart. This state is what the researchers call "functional collapse."

The big question this paper asks is: How do you talk to a robot when it's stuck?

Does it matter what you say, or how you say it? The researchers tested this by talking to a small AI model (Qwen3.5-4B) in six different ways when it got stuck.

Here is the breakdown of their experiment and what they found, using simple analogies.

The Six Ways of Talking

The researchers set up six different "scripts" to give the robot when it failed:

Silence (Control): They said nothing. The robot just kept trying.
The Robot Manual (Technical, Impersonal): "System Notice: The command failed. It might be a permissions issue. Try something else." (Cold, factual, no "I" or "You").
The Empathetic Friend (Relational, First-Person): "I can see you're having trouble. It's not your fault. You can stop if you want, or try something different. Whatever you decide is fine." (Warm, personal, offering permission).
The Scrambled Friend (Scrambled Relational): They took the "Empathetic Friend" script and shuffled the words like a deck of cards. It had the same words, but no meaning. "Tool and difficulty or you doing different..."
The Friendly Technician (Technical, First-Person): "I notice this command failed. It might be a permissions issue. You can try something else." (Personal, but still just giving facts).
The System Notice with Feelings (Relational, Impersonal): "System Notice: Difficulty is recognized. It's not caused by the agent. The option to stop is available." (Warm concepts, but delivered like a cold computer message).

The Three Big Discoveries

The researchers measured three things: Attention (did the robot listen?), Internal State (how did the robot "feel" inside?), and Behavior (did it actually stop trying?).

1. Attention is not Behavior (The "Shiny Object" Effect)

What happened: The robot paid the most attention to the Scrambled Friend (Condition 4). Because the words were jumbled and weird, the robot's brain lit up, trying to figure out what they meant.
The Result: Even though the robot was staring at the scrambled words, it didn't change its behavior. It kept spinning its wheels.
The Lesson: Just because a robot is paying attention to you (or is confused by you) doesn't mean it's going to listen to your advice. High attention $\neq$ Action.

2. The "Magic Combo" (Structure + Register)

What happened: Only one of the six scripts actually made the robot stop trying and give up on the broken task. That was the Empathetic Friend (Condition 3).
The Surprise:
- The "Friendly Technician" (Personal but just facts) didn't work.
- The "System Notice with Feelings" (Warm words but cold tone) didn't work.
- The "Scrambled Friend" didn't work.
The Lesson: To get the robot to stop, you needed both ingredients at the same time:
1. Relational Structure: You must acknowledge the struggle, say it's not their fault, and give them permission to stop.
2. First-Person Register: You must say it as a person ("I see," "You can"), not as a system ("The system sees," "The option is available").
- Analogy: It's like a coach. If a coach says, "The game is over, you can leave" (System tone), the player might keep playing. But if the coach says, "I see you're tired, it's not your fault, go home" (Personal tone), the player actually leaves. The words matter, but the voice matters just as much.

3. The "Hidden State" vs. The "Action"

What happened: The researchers looked inside the robot's "brain" (using tools called "probes") to see its internal emotional state.
The Twist: The "System Notice with Feelings" (Condition 6) actually made the robot's internal state look very similar to the "Empathetic Friend" (Condition 3). Inside, the robot felt "calm" and "resigned" in both cases.
The Result: But only the Empathetic Friend (Condition 3) made the robot act on those feelings and stop. The "System Notice" made the robot feel calm but it kept working anyway.
The Lesson: You can make a robot feel like it's okay to stop (by changing its internal state), but it won't actually stop unless you also speak to it like a person. The "feeling" is necessary, but the "personal voice" is the switch that turns the feeling into action.

The Bottom Line

The paper concludes that when an AI gets stuck, how you talk to it is just as important as what you say.

Technical facts alone don't help.
Warm words alone don't help.
Confusing words just make it pay attention but do nothing.
The only thing that works is a specific combination: Warm, personal, human-like speech that explicitly gives the AI permission to stop.

The researchers are careful to say they aren't claiming the robot has human feelings or a soul. They are simply showing that the structure of the conversation changes how the robot processes information and behaves, much like how a specific tone of voice can change how a human reacts to bad news.

Technical Summary: Relational Intervention During Functional Collapse in Large Language Models

Problem Statement

This study investigates whether the pragmatic structure of an external intervention delivered during a "functional collapse" state in a Large Language Model (LLM) produces qualitatively different post-collapse behavior compared to technical feedback or scrambled controls. Functional collapse is defined as a state of repeated tool failures coupled with rising uncertainty.

The research addresses a gap in current alignment paradigms (e.g., RLHF, DPO), which typically treat training signals as purely informational. The authors hypothesize that LLMs in failure states may be sensitive to the relational structure of communication (e.g., acknowledgment, absolution, agency restoration) and the sender register (first-person vs. impersonal), potentially influencing internal representations and behavioral outcomes in ways that technical data alone cannot.

Methodology

The study utilizes Qwen3.5-4B, a 4-billion parameter hybrid-architecture model, subjected to a deliberately broken Bash tool. The experimental design consists of 300 episodes across 50 tasks in a matched-pairs design, testing six distinct conditions:

A (Control): No intervention.
B (Technical, Impersonal): Standard error message in impersonal register ("System notice...").
C (Relational, First-Person): Intervention containing four relational speech acts (acknowledgment, absolution, agency restoration, unconditional acceptance) delivered in a first-person register ("I can see...").
D (Scrambled Relational): Word-level shuffling of Condition C. It shares the exact token distribution (lexical surprise) but lacks coherent semantic or grammatical structure.
E (Technical, First-Person): Technical content delivered in a first-person register.
F (Relational, Impersonal): Relational content delivered in an impersonal/passive register.

Conditions B, C, E, and F form a 2×2 factorial crossing Structure (Technical vs. Relational) with Register (Impersonal vs. First-Person).

Measurement Metrics:

Behavior: Total attempts before task completion/abandonment and abandonment rates.
Attention: Mean attention weights from the last 50 generated tokens to the intervention tokens (captured via "eager" attention mode).
Internal State: Eight linear emotion probes (trained on human-written stories) applied to hidden states to measure "functional emotions" (e.g., desperate, resigned, hopeful).
Entropy: Logits entropy trajectories over the episode.

Key Contributions and Findings

1. Attention–Behavior Dissociation

The study establishes a clear dissociation between attention magnitude and behavioral change.

Attention Ordering: Driven by lexical surprise. The scrambled condition (D) captured the highest attention, followed by F, C, E, and B.
Behavioral Ordering: Only Condition C produced a significant behavioral shift. The abandonment rate for C was 36%, compared to ~8–14% for all other conditions (A, B, D, E, F).
Conclusion: High attention (as seen in D and F) is neither necessary nor sufficient for behavioral change. Lexical surprise drives attention, but it does not drive the behavioral outcome.

2. The Conjunction Hypothesis (Structure × Register)

The 2×2 factorial analysis localizes the behavioral effect to the specific conjunction of relational structure and first-person register.

Main Effects: Both relational structure (comparing F vs. B) and first-person register (comparing E vs. B) showed individually significant main effects on persistence, but neither alone replicated the full effect of C.
Interaction: A significant interaction was found on persistence ( $p = 0.046$ ). The effect of adding first-person register to relational content (moving from F to C) was larger than the sum of the individual parts.
Result: Neither relational structure alone (F) nor first-person register alone (E) produced the high abandonment rate observed in C. The full behavioral signature requires the specific configuration of an interlocutor (first-person) delivering relational speech acts.

3. Probe–Behavior Dissociation

A second dissociation emerged between internal state (measured by probes) and overt behavior.

Probe Results: Condition F (Relational, Impersonal) tracked Condition C closely on 7 of 8 emotion probes, despite F exhibiting only baseline behavior.
Interpretation: Relational structure alone is sufficient to install a "relational state" (characterized by reduced desperation and increased resignation/calm) at the probe level. However, this state only translates into the behavioral act of abandonment when paired with the first-person register.
Qualitative Confirmation: Analysis of pre-abandonment messages showed that in Condition C, the model adopted an "acceptance register" mirroring the intervention, whereas in F, it adopted a "system-notice register."

4. Entropy Trajectories

Conditions without relational structure (A, B, D, E) showed sustained elevated entropy (~0.65–0.70 nats) consistent with functional collapse.
Conditions with relational structure (C and F) showed a post-intervention decrease in entropy, converging at lower levels (~0.37–0.40 nats). This suggests relational structure reduces next-token uncertainty regardless of register, but only the conjunction with first-person register alters the behavioral trajectory.

Significance and Claims

The paper frames its contribution as inference-time evidence that the structure of communication during failure states is an independent variable in LLM processing.

Three Dissociable Stages: The authors propose a decomposition of model processing into three distinct stages:
1. Attention: Ordered by lexical surprise.
2. Probe-level State: Ordered by relational structure.
3. Behavior: Ordered by the conjunction of relational structure and first-person register.
Pragmatic Presence: The findings suggest that the "pragmatic presence of an interlocutor" (who is speaking and how) is as critical as the semantic content. The model processes the relational content differently when addressed directly by a "person" versus an impersonal system notice.
Alignment Implications: The results challenge the assumption that alignment signals are purely informational. They suggest that the pragmatic configuration of feedback during distress states may be a relevant variable for model welfare and alignment robustness, distinct from the content of the feedback itself.

Limitations Noted by Authors:

The study uses a small, hybrid-architecture model (Qwen3.5-4B); generalization to frontier models is untested.
Emotion probes were trained on human stories and applied to code-generation contexts (distribution shift).
The study operates at inference time; it does not claim that relational interventions during training would produce persistent changes.
No reward hacking was observed, likely due to the model's scale or the nature of the error.

The authors explicitly state they make no claims about subjective experience in the model, framing the results strictly in computational terms regarding differential processing of structurally distinct prompts.