acceptodds
Under review as a conference paper at ICLR 2027

A Hidden Decoding Risk Makes Bigger Language Models Less Reliable

Abstract

Bigger language models are less reliable. Across three families, three benchmarks and 11 models from 0.6B to 32B, including in-the-wild chat logs, scaling closes the start-of-response knowledge gap up to 7x while within-response knowledge degradation grows up to 39x. We trace that residual to one variable, the per-position disagreement delta = log p_M - log p_O against a stronger oracle, whose second moment splits exactly into bias^2 KL(p_M || p_O)^2 and decoding risk Var[delta]. That split is an interpretability statement before it is a statistical one: the model's self-readable uncertainty H(p_M) enters only the bias term, so the risk term has no model-readable component. Risk also takes a growing share of the squared error with scale, 31% to 49% from 1.7B to 14B. At a fabrication H(p_M) relaxes within one token while risk persists for 3.4-4.6 tokens, leaving a confident-but-precarious regime that bridges consecutive fabrications (+69% at 14B). Contracting that risk at fixed KL removes 35-74% of web-verified hallucinations across six models and three families. Semantic entropy fires less on that branch (p<10^-3) though it carries nearly 4x the fabrications. Bigger models snowball mistakes faster, through a failure mode that is dominant, self-perpetuating, causal and invisible to the model itself.

Primary area:
foundation or frontier models, including LLMs
Keywords:
hallucination, factuality, large language models, scaling, long-form generation, error compounding, uncertainty quantification, semantic entropy

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_392zOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper decomposes model–oracle log-ratio disagreement into mean and variance, studies their behavior around hallucination onset, and uses oracle-guided decoding to reduce downstream hallucinations.

Critical strengths

The empirical distinction between short-lived entropy spikes and persistent oracle-referenced disagreement is investigated across multiple model families.

Critical weaknesses
  1. The “fixed-KL” causal interpretation is incorrect. Equation 4 preserves (E_q[\delta]), where (\delta=\log(p_M/p_O)), rather than the KL of the intervened distribution. Under the stated constraint, [ \mathrm{KL}(q|p_O)=\mathrm{KL}(q|p_M)+\mathrm{KL}(p_M|p_O). ] Thus the intervention does not isolate variance at fixed KL for the actual sampling distribution.

  2. Prediction 4 overstates what follows from the decomposition. Inability to recover an oracle-dependent quantity does not imply inability to predict hallucinations. Results for selected entropy and consistency detectors cannot establish the claimed limitation of all (p_M)-only detectors.

Questions for authors

What is the actual (\mathrm{KL}(q|p_O)) after intervention? What establishes detector blindness beyond the specific proxies evaluated?

Reviewer_qtz7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper evaluates the hypothesis that, "Bigger language models are less reliable. " It evaluates 3 families, 3 benchmarks and 11 models from 0.6B to 32B. The paper claims that it traces that residual to one variable, the per-position disagreement against a stronger oracle, and decoding risk. Running a AI models the paper concludes that, "Bigger models snowball mistakes faster, through a failure mode that is dominant, self-perpetuating, causal and invisible to the model itself."

Critical strengths

Lots of results, but they are AI generated., AI reported, and AI evaluated.

Critical weaknesses

The very first sentence of the abstract, "Bigger language models are less reliable." However, everything reported in the paper is supersize model generated. Paper's hypothesis is correct, then supersize model generated outputs are much less reliable. The author(s) should have clearly state the role of generative models in the abstract and introduction. I had to dig into the paper to discover that the work is not worth a human's time to read and evaluate.

Questions for authors

If authors defined the problem and used AI's help to solve it, the sequence of interactions between the authors with the AI should have been reported in the appendix or as supplement section.

Why did you not indicate the role of AI at the outset? I strongly believe that the paper is AI generated.

How many of the cited papers were read by "human" authors (if there is any human author)? it seems too many of them.

Reviewer_y7b2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how hallucinations compound during LLM generation, arguing that larger models exhibit stronger relative error propagation and that oracle-guided decoding can reduce subsequent hallucinations. While the paper is interesting, its main problem is that several central claims (including the title itself) are overstated into a level that can mislead the readership.

Critical strengths

The paper addresses an important question through useful temporal analysis of hallucination propagation, with reasonably broad empirical coverage across model families, sizes, and benchmarks (though more frontier, truly large, models could have been better).

Critical weaknesses

While the paper is interesting, its main problem is that several central claims are overstated.

  1. "Bigger language models are less reliable" is substantially stronger than the evidence supports. The results show increased relative error compounding alongside improved overall factuality. Similarly, an increasing variance share does not establish increasing absolute variance. Since reliability can have multiple definitions, the paper should define its scope precisely and qualify its title's claim.
  2. The mean–variance decomposition is an algebraic identity, and hence, it does not itself establish that sampling converts variance into subsequent bias. Similarly, a risk measure's dependence on an external oracle does not imply that model-only signals cannot predict it. The semantic-entropy experiments support a limitation of that detector in the tested setting, rather than the broader claim of structural blindness across model-only detectors. Overlal, the theoretical conclusions exceed what the decomposition establishes.
  3. The intervention does not sufficiently isolate the proposed causal mechanism. Oracle-guided reweighting changes the token distribution and may improve factuality through mechanisms beyond variance reduction. Preserving the expected original model–oracle log-probability ratio also does not preserve the intervened distribution’s actual KL divergence to the oracle, making the "fixed-KL" characterization questionable. Matched controls using temperature adjustment, truncation, and alternative oracle-guided interventions would help establish whether the gains specifically arise from the proposed decoding-risk mechanism.
Questions for authors

See above weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.