Hot Topics Flashcards
Biases in LLMS
(Hagendorff et al., 2023) (Chen et al., 2024)
LLMs are subject to some of the biases and heuristics that are observed in human reasoning - some researchers have reported more refined models respond in a more rational manner, while others report that certain biases emerge or are more prominent in later models (e.g., GPT-4)
(Yax et al., 2024) (Schramowski et al., 2022)
This may be because LLMs are increasingly being fine-tuned on the basis of human feedback, or because such biases are implicitly present in the unfiltered datasets on which they are often trained.
(Stella et al., 2023)
LLMs also exhibit biases not observed in humans. Hallucinations are one example, in which LLMs fabricate information, or sources of information, and draw incorrect conclusions on the basis of this fictitious data. LLMs also have a tendency to exhibit overconfidence in their responses, correct or otherwise, possibly because they cannot assess the quality of the data on which their decisions are predicated. Fine-tuning processes that utilise human feedback may also be driving this tendency, as humans react more positively to responses in which a model appears confident.
What prevents LLMs from reasoning like humans?
(Mondorf & Plank, 2024)
LLMs have yet to match human performance when it comes to reasoning robustly, and struggle particularly with out-of-distribution tasks.
(Mitchell & Krakuer, 2023) (Zador et al., 2023)
Some have argued that the ability to reason consistently and systematically across all domains may continue to evade models trained exclusively on language. Certain components of knowledge seem almost impossible for a model trained exclusively on language to acquire, such as fully comprehending the meaning of words, like tickle, that are intrinsically linked to sensations. Physical common-sense may contain to evade artificial intelligence models, so long as they continue to be trained on disembodied text alone. They propose what they term the Embodied Turing Test, in which a model passes the test if its behaviour (whether simulated or robotic) cannot be distinguished from that of a live counterpart.
(Bisk et al., 2020)
Capabilities that are typically achieved through interacting with one’s environment will be challenging for LLMs to acquire.
(Mitchell, 2023) (Narla et al., 2018)
Successful performance on reasoning tasks by LLMs may be because the tasks on which they are being tested are sufficiently similar to those featured in their training data, or because the benchmarks being used in these evaluations are vulnerable to statistical shortcuts. An example is an AI model successfully identifying malignant skin lesions - but had taken the presence of rulers in the images to be a cue of malignancy.
(Ullman, 2023)
Minor perturbations to standard tasks results in significant declines in accuracy rates.
