The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse

1. Problem Statement

The paper addresses the Reversal Curse, a phenomenon where autoregressive Large Language Models (LLMs) trained on a fact in one direction (e.g., "A > B") fail to retrieve the logically equivalent reverse fact ("B < A").

• The Core Issue: While LLMs can generalize flexibly from in-context information, they often lack "latent generalization"—the ability to flexibly reuse knowledge stored in their weights.
• The Question: Recent work suggests that bidirectional supervision (via Masked Language Modeling or masking-based post-training) mitigates this curse. However, it is unclear how this works mechanistically:
  1. Does the model learn a single, direction-agnostic representation of the fact (a unified concept)?
  2. Or does it simply learn a second, separately indexed memory entry for the reverse query?

2. Methodology

The authors employ a combination of behavioral evaluation, ablation studies, and mechanistic analysis (representation geometry and linear probing) across four benchmarks: Simple Reversal, Nonsense Entities, Fictional Celebrity, and Semantic Structure.

2.1 Experimental Setup

• Models:
  • From Scratch: Transformer models trained on synthetic data (Simple Reversal).
  • Pretrained: BERT-Large (MLM) vs. Gemma-3 4B (Decoder-only).
• Training Objectives Compared:
  1. NTP (Next-Token Prediction): Standard causal masking (decoder-only).
  2. MLM (Masked Language Modeling): Bidirectional attention (encoder-style).
  3. NTP+Masking: A decoder-only variant where a masked context is provided before an unmasked continuation (following Pan et al., 2025). This provides implicit bidirectional supervision.

2.2 Key Analyses

1. Behavioral Evaluation: Measuring "Reversal Accuracy" (predicting the source entity given the target and relation).
2. Ablation I (Source Prediction Necessity): Modifying training to ensure the source entity ($A$) is never masked/predicted.
3. Masking Sweep: Systematically varying which tokens (Source, Relation, Target) are masked to determine the minimal signal required for success.
4. Representational Distance: Analyzing Mean Cosine Distance between forward fact representations ($A>B$) and reverse fact representations ($B<A$) in MLP hidden layers.
5. Linear Probing (Inseparability): Training a logistic regression probe to distinguish between the vector difference of a fact and its reverse ($\Delta_{rev}$) versus the difference between unrelated facts ($\Delta_{2}$).

3. Key Results

3.1 Bidirectionality Fixes the Behavior, Not the Mechanism

• Behavioral Success: Both MLM and NTP+Masking achieve high reversal accuracy (up to 100% on synthetic data), whereas standard NTP collapses to near 0%. This confirms bidirectional supervision is sufficient to fix the behavior.
• The "Source Prediction" Requirement:
  • Ablation I: If the source entity ($A$) is never masked during training, reversal accuracy drops to 0% across all datasets.
  • Conclusion: The model must explicitly be trained to predict the source entity given the target. It is not enough to simply have bidirectional attention; the loss function must target the source.

3.2 No Unified Direction-Agnostic Concept

The central finding is that reversal success does not imply a unified latent representation.

• Representational Distance (Fig. 3):
  • In both MLM and NTP+Masking, the vector representation of the reverse fact ($B<A$) remains far from the forward fact ($A>B$).
  • They do not "collapse" into a single point. Instead, they behave as distinct entries in the model's memory.
  • Nuance: NTP+Masking shows subject-centric clustering (facts with the same source are closer), while MLM shows weaker structure where reversals resemble unrelated facts.

3.3 No "Semantic Bridge" via Linear Transformation

• Linear Probing (Fig. 4): The authors tested if a simple linear transformation could map the forward representation to the reverse.
  • Result: The difference vector ($\Delta_{rev}$) is linearly indistinguishable from the difference vector between unrelated facts (in NTP+Masking) or same-source facts (in MLM).
  • Implication: There is no consistent, reusable "reverse" vector operation encoded in the geometry. The model does not compute the reverse via a transformation; it retrieves a separate entry.

3.4 Prediction Bias (Ablation III)

• The paper reveals a severe Prediction Bias: If a token is visible in the context (via attention) but never explicitly targeted by the loss function (never masked), the model fails to learn a generatable representation for it.
• Even if the model attends to $B$ to predict $A$, if $B$ is never masked, the model cannot later generate $B$ when prompted. This suggests that "learning" in current architectures is tightly coupled to being a prediction target.

4. Key Contributions

1. Mechanistic Clarification: The paper debunks the hypothesis that bidirectional training induces a "direction-agnostic" fact representation. Instead, it proves that models learn two distinct, separately indexed memory entries for forward and reverse facts.
2. Necessity of Source Prediction: It establishes that for reversal to work, the training objective must explicitly force the model to predict the source entity ($A$) given the target ($B$).
3. Geometry of Memory: It characterizes the internal geometry of these memories, showing that while they are distinct, their organization differs between MLM (entity-centric) and NTP+Masking (more random/unrelated).
4. Prediction Bias: It highlights a fundamental limitation where information processed via attention but not targeted by the loss function is effectively "invisible" for generation.

5. Significance and Conclusion

• The "Illusion": The paper argues that the success of bidirectional objectives creates an illusion of latent generalization. The model appears to understand the fact in reverse, but it is actually just retrieving a second, pre-learned entry rather than performing a logical inversion or accessing a unified concept.
• Implications for Future Work: Simply adding bidirectional supervision (e.g., masking) is a "behavioral fix" but does not solve the underlying brittleness of latent generalization.
• Recommendation: Future objectives should aim to couple the two directions explicitly (e.g., via shared keys, explicit transforms, or consistency constraints) to enforce a unified representation, rather than relying on the model to independently learn two separate entries.

In summary, the Reversal Curse is mitigated not by creating a unified concept of a fact, but by forcing the model to store and retrieve the reverse fact as a distinct, explicitly trained memory entry.