acceptodds
Under review as a conference paper at ICLR 2027

Meaning Before Words

Abstract

Predicting words and preserving meaning are different tasks. We study this gap in a finite language model where many strings express one meaning. The probability chain rule can already represent any latent semantic generator; the obstruction lies in the decisions and resource allocation built around it. For every fixed temperature below one, we construct a full-support binary model whose semantic error tends to zero, while both token-level and sequence-level temperature sampling have error tending to one. Even exact sequence maximization fails. We give a sharp finite-length bound for the local effect. Under a shared memory limit, an optimal next-token predictor can discard an available semantic bit; we compute the exact cost of protecting it. A sharp log-loss bound separates semantic guarantees from length-normalization artifacts. Finally, a meaning-first interface makes decisions independent of verbal form when its decoder is faithful. The results separate semantic decision-making from its autoregressive language realization.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Next-token prediction, Surface form competition, Decoding strategies, Semantic representation

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

This paper analyzes when token-level prediction and decoding fail to preserve semantic decisions. It presents counterexamples involving temperature sampling, length-normalized prediction loss, and finite-memory compression, and proposes a “meaning-first” interface separating semantic decisions from their verbal realization. The analysis is interesting, but the relevance of its assumptions and the feasibility of its proposed remedy remain insufficiently established. 36611_Meaning_Before_Words

Critical strengths

The paper clearly distinguishes probability modeling, decoding, and memory allocation rather than attributing all failures to autoregressive factorization. Its theoretical constructions are explicit, and the assumptions and limitations are carefully stated.

Critical weaknesses
  1. The central failure mechanisms are largely imposed by the constructions, rather than established as characteristic of language modeling.

In Theorem 3, the semantic bit is independent of the form bits, every individual form query is more probable than the semantic query, and the original context becomes inaccessible before the query arrives. These choices make semantic retention less valuable to the prediction objective by construction. Indeed, changing the query weighting to λ>1\lambda>1 makes the optimal encoder preserve semantics without changing its architecture or capacity. The result therefore characterizes a weighted compression trade-off, but does not establish why next-token learning should systematically favor form over meaning. Moreover, independence removes the possibility that retaining a semantic fact helps predict many subsequent tokens, while the abstract memory budget is not connected to the information available through a Transformer's context, hidden states, or KV cache.

The temperature construction has a similar limitation: the incorrect meaning is concentrated in one string, while the correct meaning is spread across many alternatives. This establishes a worst-case separation, but the paper does not show whether realistic competing answers exhibit sufficient differences in form dispersion to produce a comparable effect. Thus, the mathematical counterexamples are valid, but their explanatory relevance to actual language-model failures remains unsubstantiated.

  1. The proposed meaning-first interface does not resolve the operational difficulties or demonstrate an achievable improvement.

The proposed remedy assumes a known semantic partition, a planner with controlled semantic error, and a decoder that faithfully realizes its selected meaning. These assumptions specify desirable capabilities without explaining how to learn or enforce them. The resulting error bound composes assumed planner and decoder guarantees; it does not establish that the interface makes either component more accurate.

The exact autoregressive construction likewise requires semantic continuation probabilities that may involve reasoning over possible completions, which the authors acknowledge is not an efficient algorithm. No practical approximation, implemented system, or empirical evaluation is provided to show that the proposed separation improves semantic reliability under comparable information and computational resources. Consequently, the paper identifies possible failures but does not demonstrate a workable way to address them.

Questions for authors
  1. What evidence supports the query weighting, semantic independence, memory bottleneck, and form-dispersion asymmetry assumed in the constructions? Can the authors connect these conditions to measurable behavior in actual language models?

  2. Can the meaning-first interface be instantiated without access to oracle semantic partitions or continuation probabilities, and can its benefit be demonstrated against standard decoding under comparable information and computational budgets?

Reviewer_RxZ1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

If my understanding is correct, this is a theoretical paper with no experiments and no new method. It studies how temperature scaling interacts with the semantic content of a language model's output. The authors assume a known map from token sequences to semantic labels, so that many strings can express the same meaning. They show that the most probable string need not belong to the most probable meaning. They then give several formal results in toy settings. The main one is a binary-sequence model where ordinary sampling almost always gives the correct meaning, but temperature scaling (token-level or sequence-level), greedy decoding and exact MAP decoding almost always give the wrong one. The remaining results cover a bound on how much cooling can amplify a path's probability, a statement that low average token loss does not certify semantic correctness as length grows, and a memory-budget argument that a next-token-optimal predictor can discard a semantic bit. The paper closes with a "meaning-first" interface that selects a meaning and then decodes faithfully. Overall, it formalizes the relationship between token-level and semantic-level distributions.

Critical strengths

The binary example is a clean and effective toy model for illustrating the issue. It is fully explicit and has closed-form quantities. It also shows that the failure is not an artifact of token-level temperature, since sequence-level cooling and exact MAP fail too.

The results are rigorous, and the authors are careful about the assumptions under which each one holds (Table 1 and the repeated scoping remarks).

Critical weaknesses
  • Motivation. It is unclear to me what problem this paper solves. The paper presents toy language models in which temperature scaling changes the semantic distribution, but it does not show that this is an issue for real LLMs, or how large the effect would be there. The constructions are adversarial by design, and no evidence connects them to practice. In addition, temperature scaling matters less than it used to, since many recent models are used at or near temperature 1. Even when a lower temperature is wanted, one can present the model with a list of candidate answers in the prompt and have it choose, which sidesteps the issue of many phrasings sharing one meaning. The paper does not discuss why these alternatives are insufficient.

  • Presentation. The paper is hard to read. The introduction is confusing and it doesn't clearly state what the paper does. Many sentences do not connect to one another. Examples of confusing parts:

    • "The suggestion is useful, but a common argument for it is not. Integrating out z does not itself destroy the joint distribution. An exact next-token model can reproduce every observable law produced by a latent semantic model."
    • Table 1 is not explained by the introduction
    • "We separate three operations. Decoding changes the generated distribution. Averaging log loss changes the scale of its error certificate. Compressing the context changes the information available for later prediction. None is part of the probability chain rule, so each needs its own analysis." does not tell the reader what the three operations are for or how they relate to the main result.
    • As a minor point, the paper uses terms such as "law", "sharpening" and "cooling", which are more commonly called "distribution" and "temperature scaling" in the ML literature. This more mathematical vocabulary may make the paper less accessible to an ML audience.
  • Planner–decoder interface. The meaning-first interface in Section 6 is not clearly motivated. It is unclear why a decoder is needed once a planner is available. The planner already selects the semantic decision, and the decoder only expresses a chosen label as text. For tasks where the label is the answer (classification, multiple choice, structured outputs), the decoder is unnecessary.

  • Finally, the whole analysis assumes a known map from token sequences to semantic labels, which isn't clear how one would obtain.

Questions for authors

How would one obtain the map from token sequences to semantic labels in practice? For open-ended text, how sensitive are the conclusions to errors in this map?

Reviewer_YnK8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper is concerned with the relationship between latent meanings and surface forms. It aims to clear up some possible misconceptions about how meaning and surface forms are entangled in the procedures that LM-based systems might use to make decisions.

Critical strengths

I appreciate that the authors formalize and prove their claims, and that they are explicit about the assumptions and limitations of each result.

Critical weaknesses

I am really having trouble following the thread of ideas in this paper. I would characterize the paper as a meandering series of strawmen to knock down, with no clear storyline or conclusion. These strawmen include

  • integrating out z destroys meaning
  • small token loss certifies semantics
  • many wordings force a finite model to forget meaning

The paper's apparent depth lies in knocking down these unattributed strawmen that no one has set up. The authors don't motivate the strawmen well; the reader probably doesn't have them in mind; they don't come from any cited work; and they don't seem to come from lore or widely held beliefs in the community.

Questions for authors

I have no specific questions for the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.