Reranking Laws for Language Generation: A Communication-Theoretic Perspective

To ensure large language models (LLMs) are used safely, one must reduce their propensity to hallucinate or to generate unacceptable answers. A simple and often used strategy is to first let the LLM generate multiple hypotheses and then employ a reranker to choose the best one. In this paper, we draw a parallel between this strategy and the use of redundancy to decrease the error rate in noisy communication channels. We conceptualize the generator as a sender transmitting multiple descriptions of a message through parallel noisy channels. The receiver decodes the message by ranking the (potentially corrupted) descriptions and selecting the one found to be most reliable. We provide conditions under which this protocol is asymptotically error-free (i.e., yields an acceptable answer almost surely) even in scenarios where the reranker is imperfect (governed by Mallows or Zipf-Mandelbrot models) and the channel distributions are statistically dependent. We use our framework to obtain reranking laws which we validate empirically on two real-world tasks using LLMs: text-to-code generation with DeepSeek-Coder 7B and machine translation of medical data with TowerInstruct 13B.

Paper

Similar papers

Peer review

Reviewer 8my95/10 · confidence 4/52024-07-10

Summary

This paper proposes a reranking principle for language generation from a communication-theoretic perspective. The paper conceptualizes the generator as a sender transmitting multiple descriptions of a message through parallel noisy channels. A receiver is designed to decode the message by ranking the descriptions and selecting the one found to be most reliable. Experiments show the effectiveness of proposed method in text-to-code generation task and machine translation of medical data task.

Strengths

1 This paper proposes a reranking principle for language generation from a communication-theoretic perspective. The motivation is interesting and the theoretic analysis seems reasonable. 2 The paper is well-written and easily readable.

Weaknesses

1 The related work analysis is not comprehensive. There are several ranking and reranking works in recommended systems, and none of them is mentioned or compared in this paper. 2 The experiments are not convincing enough. This paper only conducts two downstream experiments, i.e., Code generation and Machine translation. The results should be evaluated through more common and popular downstream tasks, such as QA (question choice) scenarios. 3 There is only 1 baseline for the text-to-code generation task and 2 baselines for the machine translation task. In particular, the one baseline for the text-to-code generation task is majority voting, which is not representative enough.

Questions

Please see the weakness.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer aDZG7/10 · confidence 3/52024-07-12

Summary

A number of recent works in language generation can be framed as proposing two step methods, with a method to generate proposal strings, and another to rank these strings before choosing the best one to be output (this includes, e.g., MBR decoding). This paper analyses this practice with a communication-theoretic approach. They first assume that, given a query $q$, the goal of decoding is to generate a string $y$ in a set $\mathcal{X}(q)$. They then assume the generator’s output $p(y_{1:N} \mid q)$ can be decomposed as $p(x_{1:N} \mid q)p(y_{1:N} \mid x_{1:N})$, with $x_n \in\mathcal{X}(q)$ and $p(y_{1:N} \mid x_{1:N})$ representing some kind of noise perturbation. They then analyse the probability $p(rank(y_{1:N}) \notin\mathcal{X}(q) \mid q)$ under different assumptions, showing that for many this value goes to zero as $N$ goes to infinity. Finally, they run two experiments showing their theoretically derived predictions $p(rank(y_{1:N}) \notin\mathcal{X}(q) \mid q)$ seem to correlate with the empirical probability of decoding errors.

Strengths

The paper is well written and in general easy to follow (although I think section 4 could use a bit more hand-holding). The paper provides an interesting theoretical analysis to a widely popular text generation framework. The paper then investigates whether these theoretical insights are reflected empirically in real decoding settings.

Weaknesses

In general, I liked this paper. In my opinion however, its main weaknesses are: * limited empirical evaluation, with only two tasks, one generator model, and two reranking methods (besides an oracle ranker). * the evaluation also makes some (in my opinion) debatable claims. E.g., in line 236 the authors state “[...] the imperfect reranker with majority voting, which fits the data well, as shown by the red curve.”. However, analysing Fig 4 (top), I would argue that the solid lines do not capture the data behaviour that well. In fact, the model’s performance seems to be empirically close to convergence with N, but the solid lines go monotonically down. Maybe running this analysis for larger values of $N$ would show whether the data indeed fits the predictions (specially if the predicted power law would generalise to larger values of $N$ as fit in the current data, and without fitting it on the new results).

Questions

In the code generation experiments, the paper says “we use only one test case for each problem (Shi et al., 2022), and select one candidate by taking a majority vote over the execution results, dismissing hypotheses that fail to execute on the test case.” If a single test case is used, how is majority voting performed exactly? More details here could be helpful. As a minor suggestion: I found the use of a “communication theoretical” framing here a bit distracting, and it seems to me it could be discarded with no significant change to the paper’s contributions. The authors, for instance, discuss error correcting codes early in the paper, but then they (admittedly) do not require generated strings to be error-corrected. (The selected string simply needs to be in an acceptable set $rank(y_{1:N}) \notin\mathcal{X}(q) $.) Besides, the generator and ranker are framed as a sender and a receiver—with a noisy channel in between them—but no message is actually decoded by the receiving ranker. Alternatively, highlighting the role that a communication-theoretical framing has in the paper (and why it is needed) could be useful.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors properly discuss their analysis limitations in the paper.

Reviewer aDZG2024-08-14

Response to Authors

I thank the authors for their response. I still think that running this analysis for larger values of $N$ would be good (or improving how the paper assesses that the data fits its predictions), but I have increased my scores due to the extra experiments added. I think this is an interesting paper that should be accepted.

Reviewer o8DD8/10 · confidence 4/52024-07-12

Summary

This paper proposes to regard generator-reranker LLMs, i.e., LLMs generating multiple outputs and then reranking them, as communication systems. The idea is to consider the outputs noisy with the objective for the reranker to find the less noisy one.

Strengths

- the approach is very flexible. It doesn’t depend on a particular architecture and the outputs to rerank can be generated by multiple different models. - a sound parallel is made with communication theory which helps to understand why this approach works. - the approach is well formalized. I couldn’t find any error but be aware that I’m not very familiar with Zipf-Mandelbrot and Mallows model. - Two scenarios are taken into account: with and without an independence assumption - Experiments with machine translation are very relevant for generator-reranker LLMs

Weaknesses

- Absence of analysis of the experiment results. This is a critical weakness of the paper. The experimental settings are described, and the results are given (plots), but without any comment. For the MT experiments, the authors wrote that they got some scores, and then we have the next section.

Questions

Please comment on your results. What do you conclude from them? Why are they insightful (or not insightful)? How can they be used in future work? etc. The paper is really good but we miss some analysis.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

Limitations are adequately addressed.

Reviewer EFsu7/10 · confidence 4/52024-07-12

Summary

The paper provides a framework for understanding the theoretical properties of generator-reranker systems for language generation. It relates the reranking process to error correction during the decoding of messages in noisy channels, a concept that has been well-studied in communication theory. Explicitly, the paper conceptualizes an LLM generator as a sender transmitting messages through noisy channels, with the reranker acting as the receiver decoding these messages. This framing explains why 1) things like redundancy in the set of generated strings are helpful in generator-reranker systems, actually increasing the likelihood of an acceptable output and 2) increasing the number of options from the generator in the reranking process generally increases system performance. The paper makes several theoretical contributions, showing that when generator-reranker systems meet certain theoretical requirements, there is a guarantee of “error-free” performance. This property holds even when channel distributions are not independent, i.e., when the same model is used to generate possible solutions to the input. The paper provides some empirical verification of their proposed laws.

Strengths

The paper is very well written. The math is clearly explained and sound; it provides a nice theoretical justification of why generator-reranker systems work well The topic is also very relevant, since LLM reliability (which is improved by the generator-reranker paradigm) is of utmost concern. The laws proposed in this paper also have practical use: they would allow practitioners to decide the number of strings needed from the generator system to achieve a certain accuracy, without lots of trial and error.

Weaknesses

The applicability of the communication system framework to generator-reranker systems is somewhat questionable given that there is not the same binary notion of acceptable/unacceptable for language generation systems. Rather, we’re dealing with a continuous spectrum of quality and the appeal of the generator-reranker system is its ability to increase quality (perhaps amongst “acceptable” solutions) rather than move from the realm of unacceptable to acceptable answers. The impact of this difference between the theoretical framework and the evaluation of generation systems in practice isn’t really discussed. There are a few points that could be addressed to improve readability: * Some aspects of the abstract/intro are confusing because terms have not been defined and their equivalences in an LLM generator-reranker system have not yet been specified. For example, the reader won’t know what the implications of “channel distributions being statistically dependent” are (mentioned in the abstract) until after they’ve read through much of the paper. * There isn’t much intuition about what the R.V. X corresponds to in the generator-reranker system * Some more intuition behind the scale parameter (other than just the settings that it gives for its extreme values) would be helpful The computational experiments are not very comprehensive, exploring only two generation tasks (one generator model for each task). It is thus unclear how general their results are.

Questions

* I didn't understand the (bolded) comment in lines 148-9. It makes it sound as though the quality of the reranker depends on the quality of the generator. Could this be clarified? * Minor style recommendation: Perhaps move the first sentence of 3.3 to right after providing the expression for the partition function. That feels like a more natural place to me. In footnote 1: equivalent class -> equivalence class

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors discuss most of the limitations present in their work. I would like to see an additional discussion of the implications of their results for continuous evaluation metrics (rather than binary acceptable/unacceptable)

Authorsrebuttal2024-08-06

> “I didn't understand (...) lines 148-9 (...) the quality of the reranker depends on the quality of the generator.” Thank you for letting us know that you found this part unclear. We want to clarify that the quality of the reranker itself is independent of the quality of the generator. While both perfect and Mallows rerankers achieve exponentially decreasing error probabilities as the number of hypotheses $N$ increases, the exact rate of convergence is different. For the Mallows reranker, the rate of convergence also depends on the parameters of the Mallows model ($\lambda$). As shown by Eq. (2), $P_\mathrm{err}(N; q)$ decays exponentially as $\epsilon^N$, where $\epsilon$ is the probability of generating an unacceptable hypothesis. For a Mallows model, Proposition 1 shows that the error probability also decays exponentially, with $P_\mathrm{err}(N; q) = \mathcal{O}((e^{-\lambda}(1-\epsilon) + \epsilon)^N)$. While the convergence rate $e^{-\lambda}(1-\epsilon) + \epsilon$ is generally greater than $\epsilon$, it still ensures an exponentially decreasing error probability as $N$ increases. Thus, what we meant with the bolded comment in lines 148-9 is that Mallows rerankers behave asymptotically like perfect rerankers but with a higher effective error probability, due to the additional factor depending on $e^{-\lambda}$. That is, asymptotically, a bad (Mallows) reranker is equivalent to a perfect reranker with a worse generator. We will clarify this in the final version. > “move the first sentence of 3.3 to right after providing the expression for the partition function.” Even though this result is not used in the previous section, we agree with the recommendation. We will update the paper accordingly, keeping a reminder in Section 3.3. > “In footnote 1: equivalent class -> equivalence class.” Thanks for pointing this out, we will fix the typo.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC