💬 NLP

Failure by Interference: Language Models Make Balanced Parentheses Errors When Faulty Mechanisms Overshadow Sound Ones

This paper reveals that language models fail at balanced parentheses tasks due to faulty mechanisms overshadowing sound ones, and introduces RASteer, a steering method that boosts accuracy to nearly 100% by amplifying reliable components without harming general coding abilities.

Original authors: Daking Rai, Samuel Miller, Kevin Moran, Ziyu Yao

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daking Rai, Samuel Miller, Kevin Moran, Ziyu Yao

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 you have a team of 100 experts working together to solve a puzzle. The puzzle is simple: make sure all the parentheses in a line of code are balanced. For example, if you open a bracket (, you must close it with ).

You'd think a super-smart AI (a Large Language Model) would be perfect at this. But surprisingly, even the smartest AIs often fail, especially when the parentheses get stacked up like ((())).

This paper, titled "Failure by Interference," investigates why this happens and how to fix it without retraining the AI.

The Problem: A Noisy Room of Experts

The researchers discovered that an AI doesn't rely on just one "brain" to make a decision. Instead, it uses thousands of tiny internal parts (called attention heads and neurons) that all shout out their own suggestions at the same time.

Think of it like a crowded room where everyone is trying to vote on the next word:

The Sound Mechanisms: A few experts in the room are actually very good at counting parentheses. They know exactly when to close a bracket.
The Faulty Mechanisms: The rest of the room is noisy. Some experts are confused, some are guessing, and some are shouting the wrong answer loudly.

The Failure: The AI fails not because it lacks the smart experts, but because the noisy, confused experts are shouting louder than the smart ones. The smart advice gets drowned out by the noise. The paper calls this "Failure by Interference."

The Discovery: Finding the "Super-Experts"

The researchers looked inside seven different AI models (ranging from small to large) and found something interesting:

Most parts are specialists: Many internal parts only work well for simple tasks (like one pair of parentheses) but fail when the task gets harder (like four pairs).
Some parts are generalists: A tiny handful of parts are incredibly reliable. They work perfectly no matter how many parentheses there are.
- Example: In one model (CodeLlama-7b), a single tiny part (an attention head) was actually better at solving the puzzle than the entire model was on its own!

The Solution: RASTEER (The Volume Knob)

Since the smart parts exist but are just being ignored, the researchers invented a method called RASTEER.

Imagine the AI's internal parts are speakers in a sound system.

Before: The "confused" speakers are turned up to volume 10, and the "smart" speakers are at volume 1. The result is a mess of noise.
RASTEER: This method acts like a volume knob. It identifies the smart, reliable speakers and turns their volume up slightly. It doesn't mute the others; it just makes sure the smart voices are loud enough to be heard over the noise.

The Results:

Miraculous Fixes: For some models that were getting 0% of the answers right, turning up the volume on just the top 10 smart parts boosted their accuracy to nearly 100%.
No Side Effects: Crucially, this didn't break the AI's ability to write code for other things. In fact, for some models, their general coding skills actually improved slightly.
It Works Elsewhere: They tested this on math problems (arithmetic reasoning) and saw similar improvements (up to 20% better).

What They Didn't Do (and What to Avoid)

They didn't retrain the AI: They didn't feed the AI new data or change its brain structure. They just tweaked how it uses its existing parts during the moment of thinking.
They didn't find a "circuit": Previous research tried to map out the exact path of logic (like a circuit diagram) to fix errors. The researchers found this was too complicated and didn't work well for smaller models. Their "volume knob" approach was simpler and more effective.
They didn't claim this fixes all AI errors: This specifically targets simple syntactic tasks (like parentheses and math) where the AI has the knowledge but gets confused by internal noise.

The Big Takeaway

The paper concludes that AI models are often overqualified but underutilized. They have the right tools to solve simple problems, but those tools get buried under a mountain of internal noise. By simply "turning up the volume" on the reliable tools, we can make the AI significantly smarter at specific tasks without needing to rebuild it from scratch.

Technical Summary: Failure by Interference in Language Models

Problem Statement
Despite significant advancements in code generation capabilities driven by increased model size and training data, Large Language Models (LMs) continue to struggle with fundamental syntactic tasks, specifically generating balanced parentheses. This failure persists even in state-of-the-art models like GPT-4 and Phi-3. The core challenge addressed in this work is understanding the internal mechanisms that cause these errors. While previous research has attempted to reverse-engineer the full computational circuits of LMs, the authors argue that even simple tasks involve complex "bags" of computations, making bottom-up reverse engineering difficult. The paper posits that the failure is not due to a lack of correct mechanisms, but rather a failure of interference where faulty mechanisms overshadow sound ones.

Methodology
The study employs a top-down interpretability approach to analyze seven LMs ranging from 124M to 7B parameters (including GPT-2 variants, CodeLlama-7b, Llama2-7b, and Pythia-6.9b). The methodology involves three primary phases:

Task Decomposition and Mechanism Identification:
The balanced parentheses task is decomposed into four sub-tasks based on the number of closing parentheses ( $N=1, 2, 3, 4$ ) required, as determined by tokenizer behavior. The authors utilize the "logit lens" technique to project the activations of individual components (attention heads and Feed-Forward neurons) into the vocabulary space.
- Algorithm 1 measures task correctness by checking if a component's logit projection yields the highest value for the ground-truth token compared to distractors.
- Algorithm 2 identifies token promotion by checking if a component promotes a target token above a specific threshold.
Characterization of Mechanisms:
The authors categorize components into "sound mechanisms" (reliable, generalizable across sub-tasks) and "faulty mechanisms" (unreliable, introducing noise). They analyze the precision and recall of these components, observing that LMs rely heavily on "noisy promotion," where components activate for both correct and incorrect tokens, rather than relying on highly selective, precise mechanisms.
RASTEER (Ranking and Steering):
Motivated by the finding that errors arise from the overshadowing of sound mechanisms by faulty ones, the authors propose RASTEER. This method:
- Ranks LM components based on their reliability (using metrics like recall, precision, and F1-score) and generalizability across sub-tasks.
- Steers the model by scaling the activation of the top-ranked components (attention heads or FF neurons) by a multiplier $\alpha \in [1.1, 2.0]$ before they are added to the residual stream. This amplifies the contribution of reliable components to the final logits.

Key Contributions

Mechanistic Insight: The paper reveals that LMs do not rely on a single mechanism for syntactic tasks but rather a diverse set of components with varying reliability. Crucially, it identifies that LMs fail not because sound mechanisms are absent, but because faulty mechanisms introduce noise that overshadows the sound ones.
Noisy Promotion: The study characterizes a "noisy promotion" strategy where components often promote both correct and distractor tokens, with the correct token receiving a stronger but non-exclusive boost.
RASTEER Framework: The authors introduce a steering method that systematically identifies and amplifies reliable components without requiring the reconstruction of full causal circuits.
Circuit Baseline Comparison: The work contrasts RASTEER with a "circuit baseline" derived from mechanistic interpretability (activation patching). It demonstrates that while circuit discovery is effective for larger models, it fails to provide consistent steering strategies for smaller models, whereas RASTEER succeeds across all sizes.

Results

Balanced Parentheses Performance: RASTEER substantially improves performance on the balanced parentheses task. For several models (e.g., GPT-2 XL), accuracy on the three-paren and four-paren sub-tasks was boosted from 0% to approximately 100% by steering only the top-10 to top-60 attention heads.
Component Effectiveness: Steering attention heads proved significantly more effective than steering FF neurons. Promoting the top-ranked attention heads achieved results comparable to joint promotion, while promoting FF neurons alone yielded little to no improvement, particularly for smaller models.
Generalizability:
- Code Generation: Steering for the parentheses task did not degrade general coding capabilities on the HumanEval benchmark. In fact, Llama2-7b saw a modest improvement of 5.49%.
- Arithmetic Reasoning: The method generalized to a two-operand arithmetic reasoning task, achieving performance gains of up to 20.25% (e.g., for Pythia-6.9b in multiplication).
Circuit Baseline Failure: The circuit baseline failed to improve performance on smaller models (GPT-2 variants) and, in some cases, caused performance to decline as more heads were promoted, suggesting that components identified as causally important for intermediate steps may introduce instability when steered.

Significance and Claims
The paper claims that the persistence of syntactic errors in LMs is a result of interference between competing internal mechanisms rather than a fundamental inability to learn the task. By identifying that a small subset of highly reliable, generalizable components exists within the model, the authors demonstrate that performance can be recovered by amplifying these specific signals.

The significance of this work lies in its shift from attempting to fully reverse-engineer complex circuits to a functional approach of "steering" based on component reliability. The authors note that while RASTEER is effective, the study relies on synthetic data and assumes a simple additive motif for model computation (where the final logit is the sum of component contributions). They acknowledge that this assumption may overlook non-additive mechanisms, such as active noise suppression, which remains an open direction for future research. The work suggests that functional interpretability—focusing on high-impact components for practical intervention—is a viable and underexplored path for improving model performance without retraining.