Exposing Attention Glitches with Flip-Flop Language Modeling
This addresses the brittleness in reasoning for users of large language models, but it is incremental as it focuses on a synthetic benchmark and preliminary analyses.
The paper tackles the problem of large language models producing factual inaccuracies and reasoning errors by identifying attention glitches in Transformers, using a synthetic benchmark called flip-flop language modeling to show that these models suffer from sporadic errors in long-range dependencies, with some errors eliminated via regularization.
Why do large language models sometimes output factual inaccuracies and exhibit erroneous reasoning? The brittleness of these models, particularly when executing long chains of reasoning, currently seems to be an inevitable price to pay for their advanced capabilities of coherently synthesizing knowledge, pragmatics, and abstract thought. Towards making sense of this fundamentally unsolved problem, this work identifies and analyzes the phenomenon of attention glitches, in which the Transformer architecture's inductive biases intermittently fail to capture robust reasoning. To isolate the issue, we introduce flip-flop language modeling (FFLM), a parametric family of synthetic benchmarks designed to probe the extrapolative behavior of neural language models. This simple generative task requires a model to copy binary symbols over long-range dependencies, ignoring the tokens in between. We find that Transformer FFLMs suffer from a long tail of sporadic reasoning errors, some of which we can eliminate using various regularization techniques. Our preliminary mechanistic analyses show why the remaining errors may be very difficult to diagnose and resolve. We hypothesize that attention glitches account for (some of) the closed-domain hallucinations in natural LLMs.