Large language models often achieve strong benchmark gains without corresponding improvements in broader capability. We hypothesize that this discrepancy arises from differences in training regimes induced by data distribution. To investigate this, we design controlled data interventions that isolate distributional effects under fixed training settings. We find that benchmark-aligned data improves narrow evaluation metrics while limiting broader representational development, whereas coverage-expanding data leads to more distributed parameter adaptation and better generalization. We further introduce parameter-space diagnostics based on spectral and rank analyses, which reveal distinct structural signatures of these regimes. Similar patterns are observed across diverse open-source model families, including multimodal models as a key case study, suggesting that these effects extend beyond controlled settings. A case study on prompt repetition shows that not all data artifacts induce regime shifts. These results indicate that benchmark performance alone is insufficient to characterize model capability, and highlight the importance of data distribution in shaping learning dynamics.
Figures & tables
Figure 1: Comparison of four training conditions. Condition A: baseline under the coverage-expanding regime. Condition B: baseline with an alternative learning-rate schedule (no fast decay). Condition C: repetition-concentrated regime. Condition D: frequency-concentrated regime.
2k
20k
40k
60k
72k
Trend
A: overage-expanding baseline
60.7
81.7
79.0
79.5
78.6
Stable
B: optimization-control baseline
62.1
57.6
57.6
61.2
61.2
Stuck
C: repetition-concentrated regime
58.9
79.9
81.7
80.4
81.3
Recover
D: requency-concentrated regime
61.2
77.2
78.1
77.7
77.2
Gap
Table 1: Proportion of weight matrices with α∈[2,6] (%) across training steps. Green : ≥ 78%; yellow : 70–78%; red : < 70%.
Figure 2: Layer-wise α values in self_attn.v_proj at the final training stage (72k steps). Conditions C and D exhibit stronger layer-wise heterogeneity than the baseline settings, with persistent deviations concentrated in deeper layers.
Figure 3: Change variance in mlp.up_proj . (a) Baseline-cosine-lr and (b) large-LR-no-decay share the same U-shaped depth profile, but condition B shows 1.3× amplification and non-monotonic step ordering. Conditions C and D (not shown; see Appendix) are nearly identical to (a).
Figure 4: Delta effective rank in mlp.up_proj across training steps. (a) Baseline: moderate, balanced oscillations at convergence. (c) Repeated-bias: pronounced oscillations persist in mid-to-upper layers at 72k despite aggregate [2,6]% recovery. (d) High-freq bias: compression shifts toward mid-network layers, with tighter inter-checkpoint clustering.
Figure 5: Global alpha distributions for four representative models derived from Qwen3-4B-Base. Histograms show the WeightWatcher power-law exponent ( α ) across all weight matrices. Red dashed line: overfit boundary ( α=2 ); orange dashed line: underfit boundary ( α=6 ); [2,6] proportion annotated per panel. (a) Qwen3-4B-Base: tight peak at α≈3 – 4 with sparse outliers. (b) Qwen3-VL-4B-Instruct: near-identical global shape despite deep decoder reparametrization. (c) InternVL3.5-4B-Instruct: broader distribution with a secondary hump at α≈5 – 6 , together with an elevated [2,6] proportion (73.22%). (d) AndesVL-4B-Instruct: comparable [2,6] proportion (69.60%) but denser tail mass through α=6 – 10 , indicating spectrally degraded layers that are only weakly reflected in the aggregate metric.
Figure 6: Layer-wise alpha ( α ) of self_attn.v_proj across 36 decoder layers for four instruct models. All models share a well-conditioned trough at layers 12–18 ( α≈4 – 7 ) and a spike region at layers 24–27. Key divergences include an isolated spike at layer 23 for Qwen3-VL-4B-Instruct and systematic early-layer inflation for AndesVL-4B-Instruct ( α≈7.5 – 9 at layers 0–5, versus 6.5 – 7.5 for the other models). Qwen3-4B-Instruct (LLM-only) remains the flattest, suggesting that instruction tuning without VL data is comparatively spectrally conservative. The concentration of cross-model divergences in v_proj —with much weaker separation in mlp.up_proj (see Appendix)—suggests that the attention value projection is the clearest locus of cross-model VL adaptation differences.
Figure 7: Layer-wise λmin of mlp.up_proj across 36 decoder layers for four instruct models. All models share a monotonically rising trend from ∼ 8 (early layers) to ∼ 13 (layer 33), with a terminal spike at layer 35, but a consistent rank order separates them: Qwen3-VL-4B-Instruct sits lowest, followed by Qwen3-4B-Instruct, InternVL3.5-4B-Instruct, and AndesVL-4B-Instruct highest. The separation widens from roughly ∼ 0.5 in early layers to ∼ 1.5 beyond layer 20. This rank order is stable across instruct and thinking variants (see Appendix). Since alpha and effective feature number in mlp.up_proj are nearly identical across models (see Appendix), λmin is the main differentiating metric in the MLP pathway: lower values are consistent with broader spectral resolution, whereas higher values indicate a more compressed operating regime.
Figure 8: Delta effective rank in mlp.up_proj between instruct and thinking checkpoints for four models. Qwen3-VL-4B shows large layer-specific restructuring, with both positive and negative spikes, indicating targeted rank expansion and contraction. InternVL3.5-4B shows similarly strong oscillations. AndesVL-4B-Instruct remains near zero across all layers: thinking training changes parameter magnitudes (see Appendix) but leaves rank geometry comparatively unchanged. This spectral inertness under reasoning alignment is more consistent with structurally limited adaptation than with broad representational restructuring.
Figure 9: Relative parameter change in self_attn.v_proj measured against the shared ancestor Qwen3-4B-Base for three MLLM instruct models. Qwen3-VL-4B-Instruct exhibits a bimodal structure: layers 2–20 are modified by roughly ∼ 140–150% relative to the base, while layers 0–1 and 21–35 remain near ∼ 20%, with a sharp transition at layer 21. InternVL3.5-4B-Instruct and AndesVL-4B-Instruct remain close to ∼ 20% across all layers. This large gap suggests three broad modification regimes from the shared base: minimal perturbation (InternVL base, see Appendix), conservative fine-tuning (InternVL/AndesVL instruct), and deep reparametrization (Qwen3-VL layers 2–20).
Figure 10: Weight correlation with Qwen3-4B-Base in self_attn.v_proj for three MLLM instruct models. InternVL3.5-4B-Instruct and AndesVL-4B-Instruct maintain r≈0.97 – 0.98 at every layer, preserving much of the original backbone structure. Qwen3-VL-4B-Instruct drops to near-zero correlation at layers 2–20, then recovers to ∼ 0.97 at layers 21+, mirroring the transition in Figure 9 . The simultaneous near-zero correlation and very large relative change is more consistent with deep reparametrization than with incremental fine-tuning within the modified zone.
Figure 11: Relative parameter change in self_attn.v_proj from Qwen3-4B-Base for two MLLM thinking models. Qwen3-VL-4B-Thinking shows roughly ∼ 140–150% relative change across all 36 layers, indicating that the transition at layer 21 observed in the instruct variant has disappeared. AndesVL-4B-Thinking remains near ∼ 20%, unchanged from its instruct variant. The persistent gap suggests that the modification regime is largely established during VL training and is not substantially altered by later reasoning alignment.
Figure 12: Layer-wise effective feature number of the vision encoder attention query weight ( attn.qkv_q ) across 24 layers for three MLLM instruct models. AndesVL-4B-Instruct exhibits a sharp early-layer spike at layer 3 ( ∼ 520), substantially larger than the peak values of Qwen3-VL-4B-Instruct ( ∼ 145 at layer 10) and InternVL3.5-4B-Instruct ( ∼ 275 at layer 14). At the same layer, AndesVL also shows very low α and near-zero λmin (see Appendix), consistent with a flat singular-value spectrum and poorly differentiated representations. A secondary spike at layer 10 further distinguishes AndesVL from the other models. This pattern also appears in the corresponding thinking variant, suggesting that it is established during the initial VL training stage rather than repaired by subsequent reasoning alignment.
Benchmark
Qwen3-VL-4B Instruct
InternVL3.5-4B
AndesVL-4B-Instruct
MMMU (val)
67.4
66.6
58.0
MMMU-Pro
53.2
53.5
37.6
MathVista (mini)
73.7
77.1
73.3
MathVision
51.6
54.4
27.1
MathVerse (vision-only)
46.8
61.7
34.3
DynaMath
65.3
35.7
21.2
Table 2: Benchmark comparison of 4B-scale vision-language models. * are from AndesVL’s evaluation
Figure 13: Model-level α distributions for the baseline with prompt duplication (left) and the de-duplicated condition with 75% of repeated prompts removed (right), both at checkpoint 95,000. The dashed red line marks α=2 (overfit boundary) and the dashed orange line marks α=6 (underfit boundary). The proportion of well-conditioned layers ( α∈[2,6] ) is 61.61% for the duplicated baseline and 58.93% for the de-duplicated variant. Both distributions retain similar overall shape and tail behavior, indicating that prompt repetition introduces only mild model-level spectral differences rather than a regime-level shift in global conditioning.
Figure 14: Layer-wise α values for self_attn.v_proj across 32 layers, comparing the duplicated baseline (red) and the de-duplicated variant (blue) at checkpoint 95,000. The two profiles are highly correlated overall, but localized divergences appear at layers 2, 10, and 27, where the baseline exhibits elevated α values (roughly 18–22). These selective deviations indicate that prompt duplication introduces layer-specific distortions in attention projections while leaving the global spectral structure largely intact.
Figure 15: Layer-wise effective feature number for mlp.up_proj across 32 layers, comparing the duplicated baseline (red) and the de-duplicated variant (blue) at checkpoint 95,000. The two conditions track closely throughout, with both sharing the same characteristic rising trend toward late layers. This near-invariance contrasts with the selective divergences observed in v_proj , suggesting that feed-forward layers are relatively insensitive to prompt duplication while attention projections absorb most of the structural impact.
Category
Baseline
De-duplicated
Δ
General-MLLM
0.5694
0.5892
+0.0198
General-LLM
0.3842
0.3708
− 0.0134
Business-MLLM
0.3498
0.3764
+0.0266
Business-LLM
0.4799
0.5883
+0.1084
Instruct-Following
0.5416
0.5623
+0.0208
Table 3: Benchmark performance comparison between the duplicated baseline and the de-duplicated variant across five evaluation categories. De-duplication improves four out of five categories, with the largest gains on business tasks requiring flexible interpretation. General-LLM benchmarks are largely unaffected, consistent with the intervention targeting multimodal prompt structure rather than broad language-domain knowledge.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Baseline
De-duplicated
Δ
MIA-Bench (acc)
0.7558
0.7801
+0.0243
MMMU (pass@4)
0.4756
0.5016
+0.0260
MMBench-dev (acc)
0.7147
0.7240
+0.0093
MMVet
0.4965
0.5412
+0.0448
MMStar
0.5160
0.5451
+0.0291
MathVision (avg@4)
0.1264
0.1381
+0.0117
Appendix
Table 4: General-MLLM benchmark scores for the duplicated baseline and de-duplicated variant. Scores are averaged over multiple runs where applicable.
Benchmark
Baseline
De-duplicated
Δ
MMLU-Pro (micro acc)
0.1763
0.1850
+0.0087
GPQA-Diamond (pass@4)
0.6330
0.5808
− 0.0522
CFBench (psr)
0.3433
0.3467
+0.0033
Average
0.3842
0.3708
− 0.0134
Appendix
Table 5: General-LLM benchmark scores for the duplicated baseline and de-duplicated variant.
Benchmark
Baseline
De-duplicated
Δ
OCR-GRD (avg score)
0.7185
0.8214
+0.1029
OCR-REC (score)
0.6899
0.7161
+0.0262
Screen-ViVO-Pro (avg score)
0.0724
0.0589
− 0.0135
Screen-ViVO-v2.1 (avg score)
0.1870
0.1717
− 0.0153
Object Recognition (avg)
0.3589
0.3250
− 0.0339
Info Extraction (acc)
0.2041
0.3722
+0.1681
Appendix
Table 6: Business-MLLM benchmark scores for the duplicated baseline and de-duplicated variant.
Benchmark
Baseline
De-duplicated
Δ
WritingBench (avg)
0.4813
0.5001
+0.0187
Summary (avg overall)
0.7057
0.7120
+0.0063
Scene NER (overall F1)
0.5539
0.5515
− 0.0024
Keywords (F1)
0.1786
0.5896
+0.4110
Average
0.4799
0.5883
+0.1084
Appendix
Table 7: Business-LLM benchmark scores for the duplicated baseline and de-duplicated variant.
Benchmark
Baseline
De-duplicated
Δ
IF-v2 (acc)
0.2817
0.2867
+0.0050
CFBench (csr)
0.6067
0.6133
+0.0067
CFBench (isr)
0.2500
0.2567
+0.0067
CFBench (psr)
0.3433
0.3467
+0.0033
IF-Creativity (loose acc)
0.5772
0.6296
+0.0525
IF-Info Extraction (loose acc)
0.7225
0.8197
+0.0973
Appendix
Table 8: Instruction-following benchmark scores for the duplicated baseline and de-duplicated variant.
Figure 16: Temporal evolution of model-level α distributions under the controlled data interventions. Rows correspond to the coverage-expanding baseline (Condition A), repetition-concentrated regime (Condition C), and frequency-concentrated regime (Condition D). Columns correspond to checkpoints at 2k, 20k, and 40k steps. These intermediate checkpoints make the recovery asymmetry discussed in Section 4 visually explicit: Condition C shows stronger later recovery, whereas Condition D remains more persistently offset.
Figure 17: Layer-wise α values in mlp.up_proj at the final training stage for three representative data conditions. In contrast to the stronger condition-specific separation observed in attention projections, MLP α profiles remain comparatively similar across conditions, supporting the attention–MLP asymmetry discussed in Section 4 .
Figure 18: Layer-wise λmin in self_attn.v_proj for the optimization-control baseline (Condition B). The monotonic trend provides additional evidence that this condition reflects an optimization-driven effect rather than the irregular, data-induced layer-specific distortions observed under concentrated data regimes.
Figure 19: Layer-wise effective feature number in self_attn.v_proj for the coverage-expanding baseline (Condition A) and the frequency-concentrated regime (Condition D). Condition D exhibits stronger inhomogeneity, especially in upper layers, consistent with uneven allocation of representational capacity under support collapse.
Figure 20: Change variance in mlp.up_proj for the two concentrated data regimes. Both conditions remain close to the baseline U-shaped profile reported in the main text, reinforcing the conclusion that change variance is comparatively insensitive to data regime and primarily reflects optimization schedule.
Figure 21: Additional MLP-pathway comparisons across instruct models. Both α and effective feature number show near-invariance across model families, supporting the claim in Section 5 that the MLP pathway is substantially more stable than the attention pathway.
Figure 22: Layer-wise λmin in mlp.up_proj across thinking variants. The rank ordering remains stable, indicating that the MLP-pathway separation identified in the instruct models persists under reasoning alignment.
Figure 23: Change variance in mlp.up_proj for thinking-stage adaptation. This figure supports the interpretation that AndesVL undergoes nonzero parameter updates during reasoning alignment even when delta effective rank remains nearly flat, consistent with the spectral inertness discussed in Section 5 .
Figure 24: Relative parameter change in self_attn.v_proj for InternVL at the base stage. This figure provides additional support for the minimal-perturbation regime referenced in Section 5 .
Figure 25: Additional vision encoder diagnostics for the attention query projection. AndesVL exhibits a sharp early-layer anomaly with unusually low α and near-zero λmin , supporting the interpretation in Section 5 that localized structural irregularity can appear in modality-specific submodules even when decoder modification remains comparatively conservative.
Figure 26: Delta effective rank in the attention pathway under prompt de-duplication. The duplicated baseline and de-duplicated condition remain nearly indistinguishable across layers, providing additional evidence that prompt duplication does not induce the type of regime-level restructuring observed under the controlled benchmark-shadow conditions.
Machine learning models for tabular prediction typically assume that training data are representative of the population on which they will be deployed. In practice, however, data are often collected based on convenience, availability, or cost, resulting in systematic sampling biases that can distort feature--target relationships and undermine real-world generalization. Existing tabular distribution-shift benchmarks primarily construct out-of-distribution (OOD) settings by partitioning observations across naturally occurring domains. While such partitions may induce changes in P(Y | X), the resulting shifts are neither explicitly controlled nor isolated, and the evaluation distributions themselves are often not representative of the broader target population. We introduce INCONVENIENT, a benchmark designed to study robustness to convenience-sample bias. Rather than relying on naturally occurring domain partitions, INCONVENIENT directly perturbs feature--target relationships in the training distribution while preserving the original population distribution as the deployment environment. This setting more closely reflects real-world scenarios in which models are developed from unrepresentative samples and subsequently deployed on a broader population. The distinction is particularly relevant for large language models (LLMs), whose pretraining may encode population-level knowledge that extends beyond the relationships observed in a biased training dataset. Unlike existing distribution-shift benchmarks, INCONVENIENT provides explicit control over both the direction and severity of relationship shifts, enabling a systematic assessment of robustness to convenience sampling. The benchmark comprises 200 tabular classification and regression datasets spanning diverse application domains and generates around 15,500 shifted training distributions which lead to substantial downstream degradation.
Sara Rezaeimanesh, Chad Klochko, Kundan Thind +1
Michigan State University, East Lansing, MI, USA · Henry Ford Health, Detroit, MI, USA
Representation engineering reads and steers capability directions in large language models, yet methods are typically evaluated on paper-specific synthetic data. The resulting measurements are difficult to compare or reproduce and may reflect surface patterns rather than capabilities. We present RepBench, a benchmark-grounded data layer for capability-aligned representation probing. Crawling 13,427 benchmark papers yields a taxonomy of 182 capability clusters in 13 families; harvesting 353 public benchmark datasets yields 46,149 audited probe texts covering 94 capabilities, each supported by at least two independent benchmarks. This multi-benchmark design reduces dependence on any single source: raw per-text vectors exhibit no natural cluster granularity, whereas benchmark-pooled capability vectors show an interior clustering optimum at a small number of clusters on all 12 evaluated models, with low agreement to the human taxonomy. Under cross-benchmark transfer evaluation across twelve models completed by all four readouts, difference-in-means attains the highest model-level mean on ten models, while logistic regression wins the most capability-model cells. This disagreement shows that the readout method and aggregation criterion are meaningful evaluation dimensions. The pipeline, corpus, and evaluation code are released as a reusable closed-loop workflow.
Standard evaluation protocols reveal a counterintuitive phenomenon: on 7.7% of benchmark problems spanning five datasets, larger language models underperform smaller ones by 28.4 percentage points despite 10-100x more parameters. Through systematic evaluation of 31 models (0.5B-405B parameters) across 1,485 problems, we identify the mechanism as spontaneous scale-dependent verbosity that introduces errors through overelaboration. Causal intervention experiments demonstrate this reflects correctable prompt design rather than fundamental capability limitations. Constraining large models to produce brief responses improves accuracy by 26 percentage points and reduces performance gaps by up to two-thirds. Most critically, brevity constraints completely reverse performance hierarchies on mathematical reasoning and scientific knowledge benchmarks, with large models achieving 7.7-15.9 percentage point advantages over small models -- direct inversions of the original gaps. These reversals prove large models possess superior latent capabilities that universal prompting masks. We validate findings through three independent contamination tests and demonstrate inverse scaling operates continuously across the full parameter spectrum, with dataset-specific optimal scales ranging from 0.5B to 3.0B parameters. Our results establish that maximizing large model performance requires scale-aware prompt engineering rather than universal evaluation protocols, with immediate implications for deployment: prompt adaptation simultaneously improves accuracy and reduces computational costs.
MD Azizul Hakim
Department of Computer Science Bangladesh Sweden Polytechnic Institute Chattogram, Bangladesh