The Asymmetry Between Generation and Evaluation in Language Models

28 concepts15 papers1 proposal

When humans learn a concept, our ability to produce an instance of it and our ability to recognize it are deeply intertwined. If you can solve a complex calculus problem, you can generally look at someone else's step-by-step solution and determine if it is correct. We naturally assume Large Language Models (LLMs) possess a similar, unified internal world model.

However, a striking paradox has emerged across recent literature: an LLM's generative proficiency does not imply evaluative proficiency. As explored by Marina Mancoridis et al. in Potemkin Understanding in Large Language Models, models often exhibit an internal incoherence where they fail to validate the very concepts or examples they generate. This "generator-validator gap" means that a model's outputs are often driven by probabilistic surface patterns rather than a stable, coordinated internal logic.

The Two Faces of the Disconnect

The evaluative disconnect is not a uniform deficit; rather, it manifests as a stark asymmetry that flips depending on the nature of the task.

Solving But Failing to Judge

In complex reasoning, mathematics, and open-domain question answering, models frequently generate correct answers but fail to validate them. Alice Oh et al. demonstrate this in The Generative AI Paradox in Evaluation: "What It Can Solve, It May Not Evaluate", showing that a model can successfully output a correct response but fail to score it accurately when presented with it as an external input.

Similarly, Xiang Lisa Li et al. in Benchmarking and Improving Generator-Validator Consistency of Language Models observe that a model might solve a math problem correctly, yet explicitly claim its own answer is false when prompted to verify it. This is echoed by Pardis Sadat Zahraei et al. in WSC+: Enhancing The Winograd Schema Challenge Using Tree-of-Experts, where models perform worse evaluating common-sense reasoning instances they generated themselves than they do during the generation phase.

Judging But Failing to Solve

Conversely, in domains like safety critique or complex data explanation, the asymmetry flips. Preetham Sivalingam et al. show in TSQueryBench: LLM-as-a-Judge for Time-Series Explanations that while LLMs struggle to generate numerically accurate time-series explanations from scratch, they excel at identifying and ranking correct explanations when acting as a judge. This latent evaluative capacity is also leveraged by Kai Chen et al. in Gaining Wisdom from Setbacks: Aligning Large Language Models via Mistake Analysis, where models are shown to possess an underlying ability to critique harmful content even when they are prone to generating it.

Why the Mechanisms Diverge

The root cause of this divergence lies in how different interaction paradigms shift the model's internal processing.

When generating text, a model samples tokens sequentially based on conditional next-token probabilities. When evaluating, the objective shifts to a discriminative context, altering the attention weights and prompt dynamics. Chenyang Lyu et al. in Beyond Probabilities: Unveiling the Misalignment in Evaluating Large Language Models highlight a widespread disconnect between probability-based sequence prediction (such as multiple-choice token selection) and free-form generation capabilities, proving that standard evaluation metrics often fail to capture actual generative behavior.

Furthermore, training interventions like instruction tuning can inadvertently decouple these mechanisms. As Hugh Mee Wong et al. discover in VAQUUM: Are Vague Quantifiers Grounded in Visual Data?, a vision-language model's internal probability for generating a descriptive term often directly contradicts the numerical accuracy score it assigns to that same term. This indicates that generative likelihood and explicit evaluative reasoning operate on distinct, uncoordinated internal representations.

This lack of coordination leads to logical instability. Shaobo Cui 0006 et al. explore this in Nuance Matters: Probing Epistemic Consistency in Causal Reasoning, demonstrating that a model's explicit ranking of causal factors frequently contradicts its own generated descriptions of those factors, confirming the absence of a stable underlying world model. Even worse, when acting purely as a judge, a model's explicit evaluation can be easily manipulated; Alexander Wan et al. reveal in What Evidence Do Language Models Find Convincing? that a model's explicit credibility ratings of a source do not correlate with how much that source actually influences its final answers, as they are easily swayed by superficial stylistic markers.

  ┌─────────────────────────────────────────────────────────┐
  │                   Input Prompt / Query                  │
  └────────────────────────────┬────────────────────────────┘
                               │
            ┌──────────────────┴──────────────────┐
            ▼                                     ▼
┌───────────────────────┐             ┌───────────────────────┐
│    Generative Mode    │             │  Discriminative Mode  │
│ (Next-Token Sampling) │             │ (Explicit Evaluation) │
└───────────┬───────────┘             └───────────┬───────────┘
            │                                     │
            ▼                                     ▼
┌───────────────────────┐             ┌───────────────────────┐
│ Produces a correct    │   CRASHES   │ Evaluates the exact   │
│ solution via surface  │  ◄────────► │ same solution as      │
│ statistical patterns  │  WITH...    │ incorrect             │
└───────────────────────┘             └───────────────────────┘

Reconciling Generation and Judgment

To build more reliable systems, researchers have developed methods to force alignment or find an equilibrium between these two modes.

Forcing Internal Solving and Self-Reference

If a model fails to judge what it can solve, one intuitive fix is to force it to do the work before rendering a verdict. Shaoning Sun et al. introduce this in S2J: Bridging the Gap Between Solving and Judging Ability in Generative Reward Models, showing that requiring a generative reward model to internally solve a query first drastically reduces the evaluation gap. Similarly, Wei-Hsiang Lin et al. in Do Before You Judge: Self-Reference as a Pathway to Better LLM Evaluation utilize a model's own generated answer as an explicit self-reference guide to stabilize its evaluative accuracy.

Game-Theoretic Equilibrium

Instead of choosing one mode over the other, the two modes can be pitted against each other to find a consensus. Athul Paul Jacob et al. propose this in The Consensus Game: Language Model Generation via Equilibrium Search. By treating generation and discrimination as a game and searching for a mathematical equilibrium between generative probabilities and discriminative scores, the authors achieve significantly more accurate and truthful predictions without additional training.

Explicit Alignment Fine-Tuning

Finally, models can be explicitly trained to close the gap. Juan Diego Rodriguez et al. introduce RankAlign: A Ranking View of the Generator-Validator Gap in Large Language Models, a method that directly aligns a model's generative confidence with its discriminative ranking of potential answers.

For self-improvement, Jing-Cheng Pang et al. demonstrate in Language Model Self-improvement by Reinforcement Learning Contemplation how to exploit the evaluation-generation gap in reverse, using the model's superior judging capabilities to provide a reward signal that updates its generative parameters via reinforcement learning.

Key Foundational Work

Benchmarking and Improving Generator-Validator Consistency of Language Models (2024) — contributed 3 concept elements to this direction.
S2J: Bridging the Gap Between Solving and Judging Ability in Generative Reward Models (2025) — contributed 3 concept elements to this direction.
VAQUUM: Are Vague Quantifiers Grounded in Visual Data? (2025) — contributed 3 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_evaluative_inconsistency_7dee60bf,
  title = {evaluative inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/evaluative-inconsistency/rdr_7dee60bfeba84465b664fb27498f9c8a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_evaluative_inconsistency_7dee60bf_v1,
  title = {evaluative inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/evaluative-inconsistency/rdr_7dee60bfeba84465b664fb27498f9c8a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake