We argue that there are certain properties of language that our current large language models (LLMs) don't learn. We present an empirical investigation of visual-auditory properties of language through a series of tasks, termed H-Test. This benchmark highlights a fundamental gap between human linguistic comprehension, which naturally integrates sensory experiences, and the sensory-deprived processing capabilities of LLMs. In support of our hypothesis, 1. deliberate reasoning (Chain-of-Thought), 2. few-shot examples, or 3. stronger LLM from the same model family (LLaMA 2 13B -> LLaMA 2 70B) do not trivially bring improvements in H-Test performance.
Therefore, we make a particular connection to the philosophical case of Mary, who learns about the world in a sensory-deprived environment. Our experiments show that some of the strongest proprietary LLMs stay near random chance baseline accuracy of 50%, highlighting the limitations of knowledge acquired in the absence of sensory experience.
Key Findings on H-Test
1. Insignificant Intra-Family Improvements.
A stronger model in the same model family often does not bring meaningful improvement in H-Test performance.
2. Number of Examples Has Minimal Impact.
The number of examples given neither increases nor decreases performance significantly, strongly hinting that the LLM is simply not learning from H-Test few-shot examples.
3. Deliberate Reasoning (CoT) Often Decreases Performance.
If LLMs benefit from such logical, step-by-step semantic reasoning on the H-Test, this can also imply that H-Test is fundamentally solvable by developing stronger language-only models. But CoT decreases performances in general.
4. Training with more orthography-specific language data does not improve H-Test.
We produced 1000 training instances per task in H-Test and fine-tuned gpt-3.5-turbo-0613 ten different times accordingly. After training for three epochs on each task, we evaluated them on H-Test at k = 50 and observed that no significant performance improvement was achieved.
5. Multi-modality does not automatically improve H-Test performance.
At the time of writing, LLaVA V1.6 34B is the strongest open-source multi-modal model available. Despite the addition of visual modality, we observe that simply incorporating visual data into the training does not result in a straightforward improvement in H-Test performance.
6 (Important). But the H-Test is solvable, we just don't know how.
In our paper, we have reported the seemingly unexplainable (jumping) performance improvement on H-Test from GPT-3.5 to GPT-4. This result is important as it shows that H-Test is indeed solvable (by a GPT-4-level system), but not through conventionally discussed language-only modeling techniques.
I was inspired to do this research due to examples cases given in @Owain_Evans track application for the Constellation Fellowship 2024. This research is self-funded.
Thanks for the comment. I'll get back to you sometime soon.
Before I come up with anything though, where are you getting to with your arguments? It would help me draft a better reply if I knew your ultimatum.
To consult the statistician after an experiment is finished is often merely to ask him to conduct a post mortem examination. He can perhaps say what the experiment died of.
About 1.) Agree with this duality argument.
About 2.) I'm aware of the type of tasks that suddenly increase in performance at a certain scale, but it is rather challenging to confirm assertions about the emergence of capabilities at certain model scales. If I made a claim like "it seems that emergence happens at 1TB model size like GPT-4", it would be misleading as there are too many compound variables in play. However, it would also be a false belief to claim that absolutely nothing happens at such an astronomical model size.
Our paper's stance, phrased carefully (and hopefully firmly), is that larger models from the same family (e.g., LLaMA 2 13B to LLaMA 2 70B) don't automatically lead to better H-Test performance. In terms of understanding GPT-4 performance (Analysis: We Don’t Understand GPT-4), we agreed that we should be blunt about why GPT-4 is performing so well due to too many compound variables.
As for Claude, we refrained from speculating about scale since we didn't observe its impact directly. Given the lack of transparency about model sizes from AI labs, and considering other models in our study that performed on par with Claude on benchmarks like MMLU, we can't attribute Claude's 60% accuracy solely to scale. Even if we view this accuracy as more than marginal improvement, it suggests that Claude is doing something distinct, resulting in a greater boost on H-Test compared to what one might expect from scaling effects on other benchmarks.
About 3.) Fine-tuning can indeed be effective for prompting models to memorize information. In our study, this approach served as a useful proxy for testing the models' ability to learn from orthography-specific data, without yielding substantial performance improvements on H-Test.
Abstract
We argue that there are certain properties of language that our current large language models (LLMs) don't learn. We present an empirical investigation of visual-auditory properties of language through a series of tasks, termed H-Test. This benchmark highlights a fundamental gap between human linguistic comprehension, which naturally integrates sensory experiences, and the sensory-deprived processing capabilities of LLMs. In support of our hypothesis, 1. deliberate reasoning (Chain-of-Thought), 2. few-shot examples, or 3. stronger LLM from the same model family (LLaMA 2 13B -> LLaMA 2 70B) do not trivially bring improvements in H-Test performance.
Therefore, we make a particular connection to the philosophical case of Mary, who learns about the world in a sensory-deprived environment. Our experiments show that some of the strongest proprietary LLMs stay near random chance baseline accuracy of 50%, highlighting the limitations of knowledge acquired in the absence of sensory experience.
Key Findings on H-Test
1. Insignificant Intra-Family Improvements.
2. Number of Examples Has Minimal Impact.
3. Deliberate Reasoning (CoT) Often Decreases Performance.
4. Training with more orthography-specific language data does not improve H-Test.
5. Multi-modality does not automatically improve H-Test performance.
6 (Important). But the H-Test is solvable, we just don't know how.
I was inspired to do this research due to examples cases given in @Owain_Evans track application for the Constellation Fellowship 2024. This research is self-funded.
ArXiv: https://arxiv.org/abs/2402.11349
GitHub: https://github.com/brucewlee/H-Test