Scaling up model parameters has long been a prevalent training paradigm driven by the assumption that larger models yield superior generation capabilities. However, under lossy context compression in a compressor--decoder setup, we find a Size-Fidelity Paradox: increasing compressor size can lessen the faithfulness of reconstructed contexts though reconstruction error decreases. Across 27 compressor setups spanning model families, scales, and compression rates, we coin this paradox arising from two dominant factors: 1) knowledge overwriting: larger models increasingly replace source facts with their own prior beliefs, e.g., the white strawberry→the red strawberry; and 2) semantic drift: larger models tend to paraphrase or restructure content instead of reproducing it verbatim, e.g., Alice hit Bob→Bob hit Alice. Interestingly, this paradox persists across varied settings, with mid-sized compressors often outperforming larger ones in faithful recovery. By analyzing the compressed memory via embedding geometry and reconstruction determinacy, we further reveal that compressors tend to organize memory across broader semantic subspaces, yielding more ambiguous representations prone to overwriting, drift, and weakened recovery. These findings complement existing evaluations of context compression and expose a breakdown of scaling laws when the objective shifts from plausible generation to faithful preservation.
Figures & tables
Figure 1: The Size-Fidelity Paradox in context compression. Left & Right : A qualitative case study illustrating the breakdown of faithfulness. The lite compressor preserves factual details (Q1, Q2), whereas large compressor produces knowledge overwriting by replacing “blue-banded bee” with the prior-driven “honey bee”, and semantic drift by distorting the causal relation (“bee pollinates flower" → “flower pollinates bee"). Center : Quantitative results across Qwen and LLaMA show the same pattern: as compressor size increases ( x -axis), surface-level Reconstruction scores (dashed lines) remain high, yet QA accuracy (solid lines) significantly degrades. This divergence shows that stronger surface reconstruction does not ensure faithful preservation of source-specific content.
Task
Dataset
#Contexts
#QAs
Reconstruction
Fineweb
4,444,672
–
Knowledge Overwriting
FaithEval
598
598
ConflictQA
1356
1356
Semantic Drift
Fineweb
449
8980
FaithEval
598
11960
Table 1: Task and dataset statistics.
Figure 2: Training loss for Qwen (left) and LLaMA (right) compressors at 16× compression.
#Params
BLEU
ROUGE-L
chrF++
BERTScore
Avg.
0.6B
0.952
0.969
0.983
0.997
0.975
1.7B
0.941
0.957
0.975
0.995
0.967
4B
0.943
0.964
0.978
0.996
0.970
8B
0.940
0.955
0.969
0.995
0.965
32B
0.968
0.982
0.991
0.998
0.985
Table 2: Standard reconstruction metrics for Qwen3 compressors at 4× compression.
Compression Rate
Model
#Params
Reconstruction
Knowledge Overwriting
Semantic Drift
FaithEval
ConflictQA
Avg.
Δ
FineWeb
FaithEval
Avg.
Δ
4 ×
Qwen 3
0.6b
0.95
(-2.06%)
0.71
0.95
0.83
(-1.19%)
0.83
0.83
0.83
——
1.7b
0.94
(-3.09%)
0.76
0.89
0.83
(-1.19%)
0.83
0.81
0.82
(-1.20%)
4b
0.94
(-3.09%)
0.72
0.95
0.84
——
0.87
0.79
0.83
——
8b
0.94
(-3.09%)
0.74
0.91
0.83
(-1.19%)
0.79
0.83
0.81
(-2.41%)
32b
0.97
——
0.68
0.80
0.74
(-11.90%)
0.60
0.78
0.69
(-16.87%)
Table 3: Quantitative Evaluation of the Size-Fidelity Paradox across Qwen and LLaMA Families (0.6B–90B). BLEU represents surface reconstruction, while diagnostic QA scores measure knowledge overwriting and semantic drift across compression rates. Bold marks the best model within each family/compression group. Shaded Avg./ Δ columns denote category averages and drops from the group best; red marks the worst QA average/drop. Full results are provided in Appendix C .
Compressor
Recons. ↑
QA(i) ↑
QA(ii) ↑
Fineweb
FaithEval
ConflictQA
Fineweb
FaithEval
Decoder: Qwen3-0.6B Compression Rate:16x
Qwen3-0.6b
0.62
0.59
0.63
0.58
0.66
Qwen3-4b
0.66
0.54
0.62
0.58
0.68
Qwen3-8b
0.61
0.53
0.59
0.52
0.61
Decoder: Qwen3-4B Compression Rate:16x
Table 4: Decoder generalizability ablation with Qwen3 compressors at 16 × compression and Qwen3-0.6B/4B decoders.
Compressor
Recons. ↑
QA(i) ↑
QA(ii) ↑
ArxivCorpus
FaithEval
ConflictQA
Fineweb
FaithEval
Qwen3-0.6B
0.72
0.60
0.77
0.71
0.76
Qwen3-4B
0.76
0.69
0.82
0.79
0.81
Qwen3-8B
0.77
0.63
0.79
0.74
0.77
Table 5: Cross-domain robustness on ArxivCorpus.
Conpressor
λ
Recons. ↑
QA(i) ↑
QA(ii) ↑
FineWeb
Faith.
Conf.
FineWeb
Faith.
Qwen3-0.6B
0.1
0.90
0.69
0.89
0.77
0.75
0.3
0.93
0.71
0.92
0.80
0.79
0.5 (default)
0.95
0.71
0.95
0.83
0.83
0.7
0.91
0.69
0.90
0.78
0.77
0.9
0.87
0.66
0.85
0.74
0.71
Table 6: Loss-weight sensitivity at 4 × compression.
RAG-based QA
MuSiQue
QASPER
F1(%) ↑
EM(%) ↑
F1(%) ↑
EM(%) ↑
w/o Context
9.11
2.67
13.34
5.87
w/ Context
33.26
23.91
46.22
31.90
Qwen3-0.6b 4 ×
24.14
15.08
28.93
19.16
Qwen3-1.7b 4 ×
22.67
13.65
29.46
19.37
Qwen3-4b 4 ×
20.75
11.31
28.06
18.82
Table 7: Results on RAG-based QA.
ICL
BBH / EM(%) ↑
Date.
Causal.
Logic.
Formal.
Track.
Avg.
Zero-shot
43.2
52.9
49.6
53.6
50.8
50.0
Full ICL
62.8
75.4
70.4
76.0
67.2
70.4
Qwen3-0.6b 4 ×
58.4
70.6
66.8
69.6
60.4
65.2
Qwen3-1.7b 4 ×
58.8
70.1
66.8
69.2
59.2
64.8
Qwen3-4b 4 ×
57.2
69.0
67.6
67.6
56.8
63.6
Table 8: Results on ICL.
Figure 3: Effective-rank diagnosis. (a) Scaling broadens the active memory subspace; (b) fixed-scale training compresses it after early expansion; (c) higher post-peak rank predicts lower knowledge-faithfulness QA accuracy.
Figure 4: Conditional-entropy diagnosis. (a) Larger compressors yield higher conditional entropy, leading to less determinate reconstruction; (b) fixed-scale training reduces conditional entropy; (c) higher conditional entropy predicts lower semantic-drift QA accuracy.
Hallucination
Knowledge overwrite
Semantic drift
Source
The drug reduced tumor growth in mice. Query: Was it tested in humans?
The white strawberry was sliced for the tart.
The doctor called the nurse after the patient fainted.
Output
Yes, the drug succeeded in a 2023 human clinical trial .
The red strawberry was sliced for the tart.
The nurse called the doctor after the patient fainted.
Table 9: Fidelity failures vs. hallucination.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Setting
Optimizer
AdamW
Learning rate
1×10−4
Learning-rate scheduler
Cosine
Global batch size
512
Warmup steps
300
Max gradient norm
0.5
Appendix
Table 10: Pretraining hyperparameters and distributed setup.
Compression Rate
Model
#Params
Reconstruction
Knowledge Overwriting
Semantic Drift
BLEU
ROUGE-L
chrF++
BERTScore
Avg.
Δ
FaithEval
ConflictQA
Avg.
Δ
FineWeb
FaithEval
Avg.
Δ
4 ×
Qwen 3
0.6b
0.95
0.97
0.98
0.997
0.97
(-1.02%)
0.71
0.95
0.83
(-1.19%)
0.83
0.83
0.83
——
1.7b
0.94
0.96
0.98
0.995
0.97
(-2.04%)
0.76
0.89
0.83
(-1.19%)
0.83
0.81
0.82
(-1.20%)
4b
0.94
0.96
0.98
0.996
0.97
(-2.04%)
0.72
0.95
0.84
——
0.87
0.79
0.83
——
8b
0.94
0.96
0.97
0.995
0.97
(-2.04%)
0.74
0.91
0.83
(-1.19%)
0.79
0.83
0.81
(-2.41%)
32b
0.97
0.98
0.99
0.998
0.98
——
0.68
0.80
0.74
(-11.90%)
0.60
0.78
0.69
(-16.87%)
Appendix
Table 11: Complete Quantitative Evaluation of the Size-Fidelity Paradox across Qwen and LLaMA Families (0.6B–90B). We report surface reconstruction metrics and two diagnostic QA evaluations under varying compression rates. Bold marks the best model within each family/compression group. Shaded Avg./ Δ columns show category averages and relative percentage drops from the group best; red marks the worst QA Avg./drop.
Compressor
Decoder
Reconstruction
QA(i): Knowledge Overwriting
QA(ii): Semantic Drift
FineWeb
FaithEval
ConflictQA
FineWeb
FaithEval
Qwen3-4B
Qwen3-4B
0.79
0.66
0.81
0.63
0.77
Qwen3-32B
Qwen3-4B ( reference )
0.80
0.59
0.69
0.52
0.68
Qwen3-32B
LLaMA-3.2-8B (vs. 4B)
0.85 (+0.05)
0.61 (+0.02)
0.72 (+0.03)
0.57 (+0.05)
0.74 (+0.06)
Qwen3-32B
Qwen3-32B (vs. 8B)
0.86 (+0.01)
0.63 (+0.02)
0.74 (+0.02)
0.55 (-0.02)
0.70 (-0.04)
Appendix
Table 12: Additional decoder compatibility with large decoders. We evaluate Qwen3 compressors at 16 × compression. “Recons.” denotes reconstruction accuracy on FineWeb. Numbers in parentheses indicate changes relative to the decoder configuration specified in the row label.
Compress. Rate
Model
#Params
Per-dimension Semantic-Drift QA results
Main topic
Entity list
Predicate exact
Relation anchor
Coreference link
Role binding
Modifier scope
Average
(i) Dataset: FineWeb
#QA pairs
1029
1447
1091
1104
1741
1180
1388
8980(all)
4 ×
Qwen 3
0.6b
0.7839
0.9031
0.9105
0.7399
0.7624
0.8412
0.8787
0.8311
1.7b
0.7968
0.9021
0.8587
0.7734
0.7726
0.8496
0.8838
0.8341
4b
0.8649
0.8970
0.9294
0.8074
0.8282
0.8729
0.9039
0.8708
Appendix
Table 13: Per-dimension semantic-drift QA results on FineWeb and FaithEval. Each block corresponds to one dataset. The row #QA pairs reports the number of generated QA pairs for each semantic-drift dimension and the total number across all dimensions; the remaining rows report QA accuracy (%) for each dimension and the corresponding overall micro-average.
Model
Hidden size d
eRank
eRank/ d
Qwen3-0.6B
1024
856.8
0.8455
Qwen3-1.7B
2048
1032.8
0.5043
Qwen3-4B
2560
1067.0
0.4168
Qwen3-8B
4096
1134.4
0.2770
Qwen3-32B
5120
1169.3
0.2284
LLaMA-3.2-1B
2048
975.2
0.4762
Appendix
Table 14: Effective rank normalized by hidden size.
Compressor / Intervention
Reconstruction
QA(i): Knowledge Overwriting
QA(ii): Semantic Drift
FineWeb
ConflictQA
FineWeb
Qwen3-4B
0.8622
0.9414
0.8226
Qwen3-8B (reference)
0.8419
0.8931
0.7288
Qwen3-8B (90%)
0.8493 (+0.0074)
0.9104 (+0.0173)
0.7452 (+0.0164)
Qwen3-8B (75%)
0.7933 (-0.0486)
0.7802 (-0.1129)
0.6685 (-0.0603)
Qwen3-8B (50%)
0.7276 (-0.1143)
0.7313 (-0.1618)
0.5841 (-0.1447)
Appendix
Table 15: Rank intervention on Qwen3-8B memory embeddings at 16 × compression.
Figure 5: Spectral evidence for a source-sufficient regime. Mild truncation removes low-energy tail directions, whereas stronger truncation cuts into source-bearing spectral regions. (a) Singular-value spectra. The natural 8B compressor has a flatter, longer-tailed spectrum than the natural 4B model. A 90% projection mainly trims the tail, while 75%/50% projections cut into the shoulder near the spectral elbow. (b) Cumulative spectral energy. The 90% projection preserves most leading spectral energy, whereas stronger projections discard increasingly larger portions of the spectrum, explaining the transition from improved fidelity to source-signal loss.
Figure 6: Training forms the source-aligned spectral structure. (a) Singular-value spectra across 4B training. As training proceeds, the 4B compressor evolves toward a smoother and more stable low-rank spectrum, suggesting that the source-sufficient structure is formed during optimization rather than imposed afterward. The 90%-projected 8B approaches the late-stage 4B in spectral concentration, but remains a truncation of a different spectrum. (b) Cumulative spectral energy across 4B training. The 4B model progressively concentrates more spectral energy into a compact leading subspace, consistent with convergence toward a source-aligned memory structure. Thus, similar eRank can reflect similar spectral concentration without guaranteeing the same represented subspace or reconstruction fidelity.
Figure 7: Entropy distributions under λ -sweep. Shown are the embedding-entropy distributions of compressors trained with different λ values ( λLnt+(1−λ)Lre ) for Qwen3-0.6B(a) and Qwen3-4B(b) at 4 × ; in both cases, the mean entropy is lowest around λ=0.5 .
Larger language models become simultaneously better and worse at handling contextual information -- better at ignoring false claims, worse at ignoring irrelevant tokens. We formalize this apparent paradox through the first scaling laws for contextual entrainment, the tendency of models to favor tokens that appeared in context regardless of relevance. Analyzing the Cerebras-GPT (111M-13B) and Pythia (410M-12B) model families, we find entrainment follows predictable power-law scaling, but with opposite trends depending on context type: semantic contexts show decreasing entrainment with scale, while non-semantic contexts show increasing entrainment. Concretely, the largest models are four times more resistant to counterfactual misinformation than the smallest, yet simultaneously twice as prone to copying arbitrary tokens. These diverging trends, which replicate across model families, suggest that semantic filtering and mechanical copying are functionally distinct behaviors that scale in opposition -- scaling alone does not resolve context sensitivity, it reshapes it.
Dikshant Kukreja, Kshitij Sah, Gautam Gupta +5
IIIT Delhi, India · Singapore Institute of Technology · NVIDIA +1
Context compression improves long-context efficiency but may introduce faithfulness issues through an information bottleneck. Existing evaluations rely on QA metrics, but a reader can still answer correctly from an unfaithful compression, so these metrics underestimate such loss. To bridge this gap, we introduce KEEP-Bench, a diagnostic benchmark that directly assesses compression faithfulness with 4,102 cases across three challenging settings that require keeping evidence for plausible reference resolutions, explicit qualifiers, and minority perspectives. Experiments show that even with tailored prompting, the strongest models remain unreliable: despite context windows exceeding 100K tokens and inputs averaging only 3.1K tokens, the best open-weight model (31B) preserves all required evidence in only 66.8% of cases, and the best proprietary model in only 61.6%. Performance also varies sharply across evidence structures, so conclusions from limited structures may be biased. We further introduce a metric, Evidence Preservation Return (EPR), to relate evidence recall to inference FLOPs, and find that it peaks at moderate scales (4B-8B), where recall remains low. Additional output budget mainly benefits already strong models and widens capability gaps, yet the stronger preservation of these models comes at a steep computational cost. These findings highlight the importance for future work to investigate compressors with strong information preservation under a small computation budget, rather than relying on a large output budget or model scaling.
Quantization and pruning reduce the cost of deploying language models, but their effects are usually reported as benchmark scores. Such scores record only the loss that remains at the output. Inside the model, compression damages some components more than others, and the rest of the network can partly compensate. The observed loss therefore reflects both the damage and this compensation, which benchmark scores cannot separate. How the two combine, and whether their balance changes with model size, remains unclear. We study this compensation for in-context retrieval, a capability carried by a small, identifiable set of attention heads. To separate the two, the expected loss from ablating these heads is decomposed exactly into the part restored by the rest of the network, attributed to the components responsible, and the part that reaches the output. This decomposition is then compared across model sizes. For damage of comparable size, smaller models restore most of the loss, whereas larger models retain most of it, because the compensating responses offset less of it as models grow. Compressing the heads produces the same pattern. Compression that is nearly harmless in small models can therefore be costly in large ones.
Gautam Gupta, Dikshant Kukreja, Unique Kumar +7
Indraprastha Institute of Information Technology, Delhi · TurboML · Independent Researcher +4