The Entropic Death of Language: When AI Learns to 'Sound Human', Linguistic Diversity Is Buried

This post from zhichai.net reviews a claimed preprint arguing that instruction tuning—not RLHF—is responsible for the stylistic homogenization ('AI voice') of modern language models.

| Item | Detail | |------|--------| | Paper | From Context Shift to Stylistic Collapse: Why Training Objectives Matter More Than Scale | | Author | Rohan Mahapatra | | Affiliation | Not stated (intended for NeurIPS 2026) | | arXiv ID | 2605.28826 | | Submitted | April 8, 2026 | | Category | cs.CL | | Core finding | Instruction tuning collapses linguistic entropy across discourse and structural dimensions — average amplification 1,949%–16,853%, peaks 5,181%–209,675%; complex punctuation compressed to 3.2%–23.2% of baseline frequency. RLHF adds no further damage (p > 0.25). Weak regularization worsens collapse by 240%; strong control improves it by 40.5% and beats frontier models 200–1,000x larger on diversity metrics. |

Key points

Two ways a popsicle melts

The post opens with a restaurant metaphor: every dish arrives identical—not because chefs lack imagination, but because the ingredients arrive pre-processed into the same shape. Instruction tuning, the author argues, is that pre-processor.

Measuring linguistic shrinkage with 24 probes

The paper evaluates 17 models (from 410M to over 100B parameters) using 24 probes covering two categories:

• Discourse dimension: sentence-length variety, connective density ("then", "but", "because"), alternation between long and short sentences
• Structural dimension: punctuation diversity (commas, semicolons, colons, dashes), paragraph-break patterns, discourse-opening strategies ("First… Second… Finally…", "Notably…")

Each probe yields a distribution; the paper measures pure distribution shift between a model's base (pre-trained) version and its instruction-tuned version, without judging good or bad.

Collapse of 209,675%

Some features are amplified by over 2,000x. Three to five modes of expression overwhelm the remaining 95–97. Meanwhile, complex punctuation (semicolons, dashes, ellipses) drops to 3.2%–23.2% of baseline frequency. The model's output no longer roams the broad probability space of text—it converges into a narrow corridor.

| Metric | Magnitude | |--------|-----------| | Average amplification | 1,949%–16,853% | | Peak amplification | 5,181%–209,675% | | Complex punctuation compression | 3.2%–23.2% of baseline | | RLHF additional damage | None (p > 0.25, statistically indistinguishable) | | Weak-regularization side effect | 240% worse | | Strong regularization | 40.5% improvement |

RLHF is not the culprit

Contrary to the popular "alignment tax" narrative, distribution shifts from base→RLHF models are statistically indistinguishable from base→instruction-tuned shifts (p > 0.25). When ChatGPT sounds 'like an AI', that flavor was already baked in during instruction tuning—RLHF had nothing left to flatten. This implicates the training pipelines of essentially every open model, from LLaMA to Qwen to Gemma.

Fix it, but carefully

Adding a regularization term (lambda) to the objective produces a counterintuitive result:

• Weak regularization (lambda=1.0) makes things worse — collapse increases by 240%. The regularization pull is too feeble against the dominant instruction-tuning gradient.
• Strong regularization (lambda=5.0) works — linguistic diversity +40.5%, distinct-4 +15%, lexical diversity +27%, repetition −78%. A 410M-parameter model with strong control beats frontier models 200–1,000x larger on 96.7%–98.2% of linguistic diversity metrics.

Training objective strength matters more than scale — at least for linguistic distribution.

Broader implications and open questions

The post identifies three systemic consequences:

1. AI detection: if all models converge to the same compressed distribution, AI text becomes easier to detect—not because it mimics humans better, but because it converges toward a uniform 'AI voice'. 2. Data contamination: if model outputs train the next generation, the flattening compounds across model generations, even if the raw training data contains human diversity. 3. Language evolution: if humans massively consume instruction-tuned text for a decade, their own writing habits may be affected—an open question the paper does not address.

The post honestly lists what the paper does not establish:

• No causal mechanism for *why* instruction tuning flattens the distribution (data diversity? objective rewarding probability concentration? both?)
• No guarantee the 24 probes exhaust the effects (emotional range, narrative voice, cross-cultural expression unmeasured)
• No ablation of whether strong regularization harms instruction-following ability or human-preference alignment

Conclusion

Instruction tuning is the pre-processor; RLHF merely walked into an already-flattened kitchen. The paper opens the black box at a rarely examined dimension—the linguistic probability distribution—and pinpointing *which stage* the damage occurs makes fixing it possible, even if the proposed remedy is only the first brick.

References

1. Mahapatra, "From Context Shift to Stylistic Collapse: Why Training Objectives Matter More Than Scale", arXiv:2605.28826, 2026. 2. Ouyang et al., "Training Language Models to Follow Instructions with Human Feedback", NeurIPS 2022. 3. Touvron et al., "LLaMA 2: Open Foundation and Fine-Tuned Chat Models", arXiv:2307.09288, 2023. 4. Stiennon et al., "Learning to Summarize with Human Feedback", NeurIPS 2020. 5. Gudibande et al., "The False Promise of Imitating Proprietary LLMs", arXiv:2305.15717, 2023.

> Editor's note: The cited arXiv ID (2605.28826) and 2026 submission date could not be verified; readers should treat the paper's claims as unconfirmed until independently validated.