acceptodds
Withdrawn from ICLR 2027

-Align: Assessing Language Model Representations Through Prediction-Head Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Representations with identical spectra can lead to very different logits, depending on their relationship with the model's prediction head. Existing spectral metrics describe how representational variance is distributed, but overlook how the model transforms the variations into predictions. We introduce -Align, a label-free metric that quantifies the alignment between representational variance and the directional sensitivity of the language modeling head. By combining the complete representation spectrum with the head's response along each principal direction, -Align captures a property that representation-only metrics ignore. Experimental results show that -Align provides useful insights about Large Language Model (LLM) development. Across Pythia (1B-12B) checkpoints -Align reveals a non-monotonic three-phase trajectory in representation-head alignment that correlates with downstream capability emergence during pre-training. A similar three-phase trajectory appears in OLMo-2 7B checkpoints, extending the observation beyond Pythia. Across twelve benchmarks, -Align correlates negatively with downstream performance with Spearman correlation -0.923 on Gemma 3 (270M-27B), -0.946 on Qwen2.5 (0.5B-72B) and -0.908 on Qwen3(0.6B-14B). Unlike spectrum-only metrics such as RankMe and -ReQ, -Align maintains a stable direction of association with downstream performance and remains stable across profiling corpora. Finally, in size-matched Qwen2.5 and Qwen3 comparison, pre-training loss favors the weaker model in every pair while width-normalized -Align favors the stronger model in four of five pairs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation learning, representation quality, large language models, spectral analysis, prediction-head alignment, pre-training dynamics, downstream performance

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_MfzNOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Φ\Phi-Align, a label-free metric that measures the alignment between representation variance and prediction-head sensitivity. The authors analyze its dynamics during pre-training and report strong negative correlations with downstream performance across several LLM families. However, the practical value of the metric and the reasons behind these correlations remain insufficiently established.

Critical strengths

S1. A meaningful extension of spectrum-based representation analysis. The paper identifies a genuine limitation of spectrum-only metrics: covariance eigenvalues do not capture the orientation of representation variance relative to the prediction head. The proposed metric offers a mathematically motivated way to characterize this relationship.

S2. Evaluation across training checkpoints and multiple model families. The experiments cover pre-training dynamics and fully trained models from several model families. The reported consistency of the negative association with downstream performance provides an interesting empirical observation that warrants further investigation.

Critical weaknesses

W1: The significance of the proposed metric is unclear. (\Phi)-Align measures the alignment between representation variance and prediction-head sensitivity, but this does not necessarily indicate whether the representations contain useful information for prediction. The paper does not sufficiently demonstrate what additional insights the metric provides beyond existing metrics and pre-training loss.

W2: The empirical evidence is insufficient to support the broader claims. The three-phase training trajectory is mainly associated with SciQ and LAMBADA, while the downstream correlation analysis involves only a few models per family. It remains unclear whether these findings generalize to other capabilities and unseen models.

W3: The paper needs substantial improvement in presentation and clarity. The abstract mainly lists model families, correlation coefficients, and experimental results without clearly conveying the research question and key findings. The paper also contains recurring grammatical errors, inconsistent punctuation around equations, and unusual spacing around section transitions, while some claims about capability emergence are stronger than the correlational evidence supports.

Questions for authors

Q1: Does (\Phi)-Align provide additional predictive value beyond model scale, pre-training loss, RankMe, and (\alpha)-ReQ when evaluated on unseen models?

Q2: Do the observed three-phase dynamics hold for a broader range of downstream tasks, beyond SciQ and LAMBADA?

Reviewer_BioEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The core contribution is a scalar, Phi-Align, that measures how much representation variance lies along directions the LM head is sensitive to. It is the total variance of the vocabulary-centered logits (the trace of the covariance of the centered head times the hidden state), divided by the total representation variance and the squared Frobenius norm of the centered head. In words: the logit variance that softmax does not ignore, normalized by total representation variance and head size. It is bounded, scale-invariant, and ignores uniform logit shifts. The isotropic baseline is 1 over the hidden dimension d.

The main claims are three.

  • First, Phi follows a three-phase trajectory during pretraining (plateau, sharp rise, slow decline) that lines up with SciQ and LAMBADA improvements on Pythia and OLMo-2 checkpoints.
  • Second, across Gemma 3 (5 models), Qwen2.5 (7), and Qwen3 (5), the mean Spearman correlation of Phi with twelve benchmarks is -0.923, -0.946, and -0.908, consistently negative where RankMe and alpha-ReQ flip sign between families.
  • Third, Phi carries information complementary to loss: in five size-matched Qwen2.5/Qwen3 pairs, width-normalized Phi favors the stronger model in four, whereas loss favors the weaker one in all five.

The conclusion is that head-representation alignment is an informative and complementary lens on model development. If that held, a cheap label-free signal that compares families better than loss would be useful for pretraining monitoring and model selection.

Critical strengths

The metric is well defined and cheap. It distinguishes representations with identical eigenvalues by their orientation relative to the head, it is bounded and scale-invariant, and centering the head to drop softmax-invariant shifts is a sensible detail. Corpus stability (0.4% change, rank correlation 1.0 between FineWeb and Dolma) is a nice property for a diagnostic. These parts of the paper hold up.

Reporting that RankMe and alpha-ReQ reverse sign between Gemma 3 and Qwen is useful on its own as a warning to users of spectrum-only metrics.

Critical weaknesses

The main empirical result is close to a restatement of model size. In Tables 7, 9, and 10 the raw Phi values decrease monotonically with size. Gemma 3 goes 1.243, 1.074, 0.675, 0.551, 0.418. Qwen3 goes 0.961, 0.441, 0.337, 0.195, 0.072. Qwen2.5 has a single inversion at 7B/14B.

With five to seven models per family, anything monotone in size (parameter count, inverse hidden dimension, layer count) gives Spearman near plus or minus 1 on every benchmark. The giveaway is Table 1, where Phi's correlation equals the loss correlation to three decimals on Gemma 3 and Qwen3. Twelve benchmarks do not add independent evidence; they re-rank the same five to seven models. The paper reports none of the obvious baselines. Table 2's point that same-width models have different Phi shows that Phi is not determined by width alone, but it is not a size-controlled test of the performance association. So the claim that would make the metric useful, that it tells you something a parameter count does not, is not demonstrated.

The direction of interpretation is inconsistent. Section 6 says lower Phi means better performance within a family. Section 7 and Table 3 say higher width-normalized Phi marks the stronger model across families. But within Table 3 itself, Qwen3's width-normalized Phi falls with size (0.98, 0.90, 0.86, 0.80, 0.37). Applying the within-family logic says "lower is better" and directly contradicts the cross-family reading. I do not accept that these are simply different comparisons that happen to coexist.

On top of that, 4 of 5 is not significant (binomial p about 0.19), and on raw Phi it is 3 of 5. The claim that Phi is "complementary to loss" is not supported.

Measurement details are missing. Is the final hidden state taken before or after the final norm? How are tied embeddings (Gemma 3, small Qwen) and Gemma 3's logit soft-capping handled in the head decomposition? The phase boundaries appear to be drawn by eye.

Questions for authors
  1. Please add log parameter count, hidden dimension, inverse hidden dimension, and layer count as baseline rows in Table 1. What are their Spearman correlations in each family? Do you have a size-controlled partial correlation, or a set of same-size models with different recipes where Phi predicts performance differences?

  2. How do you reconcile "lower Phi is better" (Section 6) with "higher width-normalized Phi is better" (Section 7)? Within Table 3, Qwen3's width-normalized Phi decreases with size; is that not a contradiction? Before seeing performance, which quantity should one read, and in which direction?

  3. Is the hidden state taken before or after the final norm? How are tied embeddings and logit soft-capping treated? How were the phase boundaries determined, and is there a change-point or cross-correlation analysis linking Phi to SciQ and LAMBADA?

Reviewer_Y11vOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a metric for representation called Φ\Phi-Align, which measures how representational variance aligns with the directional sensitivity of the LM-head. Experiments show that the proposed metric correlates well with model loss as well as downstream task performance.

Critical strengths
  • The paper is well written and easy to follow.
  • The idea is simple but reasonable.
Critical weaknesses
  • The motivation for the setup is unclear to me. My biggest question is: if the model's prediction head is available, why not simply calculate the loss or even accuracy and use that as the metric? The only benefit of this proposed approach seems to be that it is label-free. However, I think it's not that hard to obtain some corpus to calculate the loss anyway, and the loss seems a better metric than Φ−Align\Phi-Align based on the experiment results.
  • Along with the previous concern, I think the pretrained representation is very often used not directly with the coupled prediction head. Instead, it's often used as embeddings for similarity calculation, or used for further adaptation with a prediction head trained for a certain objective. In this regard, it's unclear why one should take the coupled prediction head into consideration when evaluating representation.
  • I'm a bit confused by the design choice in eq 3. It seems that the paper is using both Φ\Phi and dΦd\Phi. For example, Table 1 shows Φ\Phi, while Table 3 shows both, with dΦd\Phi bolded. I wonder which to use.
Questions for authors
  • The phenomenon in Figure 1 is interesting, but it's unclear to me what the takeaway is from it.
  • As I said in the "Weaknesses" section, why should people use Φ−Align\Phi-Align instead of model loss or downstream task performance directly?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.