Talk Less, Interact Better: Evaluating In-context Conversational Adaptation in Multimodal LLMs

Thank you for your comments!

I think the results are stronger than the authors let on...

We will improve and expand the discussion (if accepted, with the extra camera-ready page). The framing you suggest (agent vs next-word predictor) nails it: LLMs have no reason to communicate efficiently (no cost/effort). We hypothesized that they may still acquire it by training on human data. Our results clearly show this is not the case, which can be considered surprising given how much human data is used and how well the models imitate it.

Experiments on using prompts to induce… bit unnecessary

While we hope adaptation to arise naturally, we need to answer questions about inducing it (e.g., as raised by R-XRWU), and prompting is often considered promising. We also wanted to see if MLLMs do have the ability for Gricean behavior, and maybe just not show it without being “prompted”. Our results show this is not the case, and prompting is not sufficient for such pragmatic behaviors.

More profoundly, these variants reveal underexplored causes of failures. For example, the gain in lexical efficiency without message convergence under some prompts is a surprising finding, which will caution future researchers not to rely on length reduction as the sole indication of efficiency. Overall, to offer a framework for evaluating MLLMs, we wanted ICCA to be thorough and address the different ways people may use MLLMs (e.g., different prompt intervention/interaction formats).

Engineering prompts for each model

We wanted to give models the best chance, even at the cost of extra development. For example, a prompt that worked well for one model might lead another model to sometimes output invalid outputs (e.g. “I’m not sure about the answer”), whereas additional instructions against these trivial but adverse behaviors may hurt the models not displaying these behaviors (verbosity of the instruction can hurt LLM performance [1]). By adopting model-specific prompts, we can mitigate these issues. With this being said, all model-specific prompts under the same experiment setup are based on the same instruction. Only the prompts across different interaction variants (different experiment setups) have a substantial difference.

[1] https://arxiv.org/pdf/2307.03172

Related work

Thank you for suggesting the related works! We will discuss the photobook dataset, Chalamalasetti et al 2023, and the broad picture of using self-play to evaluate LLMs as conversational agents.

Thank you for your comments!

Comparison with Hawkins et al (2020ab)

We are inspired by Hawkins’ works, rely on their data, and discuss them extensively. We will enhance the discussion of how our work differs. First, we take their data and the reference game scenario in a very different direction, both our object of study and research questions are different. Hawkins presents an approach of adapting through tuning model parameters during an interaction. We focus on diagnosing the behavior of LLMs (a class of models very different from the models Hawkins studied) and on in-context learning/adaptation. Additionally, our evaluation differs. Hawkins focuses only on the standard setup. When a model fails under the standard setup, there can be many contributing factors to the failure. We design variants specifically to diagnose the types and causes of model failures for in-context LLM adaptation. As you suggest, we will communicate these differences more clearly and explicitly.

Adding related works

These are great suggestions! We will improve the paper to discuss [1-4] (if the paper is accepted, we get an extra page for the camera ready).

It’s particularly relevant that [3] shows the unintended consequences of using explicit prompts to help form common grounds, which is consistent with our findings and intuition that explicit prompt intervention cannot solve adaptation failures. Discussing [3] will provide a more comprehensive picture of this issue. The synthesis of our work with [3] and [4] also highlights interesting connections. Both [3] and [4] study scenarios where common ground requires explicit acts (e.g., clarification/followup questions) and LLMs presume such common grounds too easily without adequately using those acts. In our case, ad-hoc conventions form without explicit actions in humans, whereas LLMs fail to do so. These different types of common grounds bring about different challenges. This highlights both the richness and the importance of the space.

Anthropomorphizing language should be avoided

We completely agree with the importance of avoiding such language! We will revise the writing.

Thank you! We are glad you appreciated our novelty, clarity, and soundness.

Significance of the results

Are our results surprising? The progress of LLMs has been impressive, with new results often surpassing human performance. LLMs are trained on human language data, which is rich in adaptation and convention formation/use. Given these models’ success, one would expect them to acquire these characteristics of human language as well. Our results show the opposite: while some aspects of language are well-addressed by current paradigms, others are not. Beyond the fundamental learning problem, our work’s importance was made clear by the recent conversational use cases LLM companies have been demoing (e.g., OpenAI’s latest demo). Overall, this illustrates that the problem is both important and underdiagnosed by current research. ICCA is a framework to drive important future development, in a way that was not previously on the agenda or easy to diagnose. We will improve the presentation of the significance in the paper.

Breadth & depth of the results

Indeed, we observe MLLMs do not display in-context conversational adaptation. Our analysis and ICCA itself go deeper though, including:

• Even with explicit instruction on length reduction (Variant S3), the models display lexical efficiency but miss critical aspects of true communication efficiency (convergence and consistency). This finding is unexpected (contrary to human behaviors). It emphasizes that communication efficiency is not only about length reduction. ICCA gives future researchers the tools to look beyond length.
• Some models can display seemingly good listener accuracy by memorizing textual information without forming conventions on visual signals. Our different interaction variants make ICCA an effective tool for such diagnosis. This is not possible using the standard evaluation from the prior work. We discuss this in Appendix D, but will expand the discussion in the paper.
• Associative recall, the ability to recall information mentioned in-context is believed to be predictive of in-context learning quality [1]. We observe underexplored implications of this ability. Throughout an interaction, IDEFICS and LLaVa assign high probability to the exact repetition of initially used phrases, marking a good associative recall. But this hinders the development of more concise phrases. We will further elaborate on this issue when we revise the paper.

[1] https://arxiv.org/pdf/2312.04927

Thank you for your comments!

the term adaption is way too vague to describe the phenomenon taking place during reference games.

We use “adaptation” following established conventions in cognitive science that describe the phenomena humans show in reference games: spontaneously and gradually adopting more efficient language and understanding the interlocutor better. These are exactly the behaviors we are interested in. Our abstract and introduction clarify that our goal is to create tools to evaluate whether MLLMs display such adaptations as humans do, which are essential for communication efficiency. We will make sure the intro relays the term "adaptation" in a concrete and clear way.

Also, such adaptations are broad phenomena, and not just about reference games. But, they are known to be well-demonstrated, measurable, and extensively studied in cognitive science through reference games. If LLMs acquire linguistic abilities similar (or better) to humans, reference games make for an ideal testbed.

And even more it can let think that the LLMs may be adapted to reach such specific objectives.

The title specifically focuses on “in-context” behaviors of MLLMs, so it clearly states the kind of adaptation we are studying, meaning: in contrast to tuning parameters or changing the architectures. In-context learning is well established in the study of LLM as a way to modify (i.e., adapt) LLM behavior. We also discuss the relatedness of in-context adaptation and few-shot in-context learning in the related work section. We will further clarify this distinction as early in the paper as possible.

one may straightforwardly think of means to accompany these prompts and re-enforce their effects...

Our focus is evaluating if LLMs without further modifications display this behavior, to see if it arises from their training. This is key to displaying this behavior in a truly general way. Our goal here is not just to understand the abilities of LLMs, but to contribute to the fundamental understanding of what arises (and what does not) from the way LLMs are trained. This is why we made prompting efforts specifically to elicit behavior that is already baked into the model, but maybe doesn’t show for one reason or another. Building on top of LLMs to further address their deficiency in adaptation is a good direction for future work, one that is enabled and supported by the ICCA framework.