Meet PLASMA: A Novel Two-Pronged AI Approach to Empowering Small Language Models with Procedural Knowledge and Planning Capabilities (Counterfactuals)

Large Language Models (LLMs) excel at many downstream tasks that require common sense, thanks to their large size. One such activity is procedural planning, which involves breaking down a high-level goal into a logical, compelling, goal-oriented series of actions (plan) (e.g., “Watch a movie,” “Find movie showings,” “Choose a movie”,…). Recent methodologies use LLMs to model this job as a conditional text generation problem. LLMs work well in the job, but widespread implementation of LLMs is hampered by their high computational cost and accessibility problems.

Researchers at the Allen Institute for Artificial Intelligence, the University of Washington, the University of Southern California, the University of Tohoku, and the University of Pittsburgh provide PLASMA (PLAn with Tiny Models), a cutting-edge two-pronged framework to help tiny LM to acquire planning. skills. They use an inference time decoding technique to enable structured reasoning and distillation of symbolic procedural knowledge to improve implicit knowledge in small LMs (Figure 1). They propose a two-stage formulation of distillation of extended procedural knowledge:

(i) verbalization of knowledge to produce procedural knowledge from an LLM and

(ii) knowledge distillation to transfer the knowledge produced by the LLM to a smaller LM.

They verbalize information for innovative task formulations in counterfactual circumstances, such as counterfactual planning and review, in addition to the traditional planning task.

Figure 1: Distillation of knowledge from symbolic procedures

In particular, the model develops or modifies a plan based on a specific goal (eg, “watching a movie”) while adhering to an additional constraint (eg, “at home”). These tasks provide a more realistic environment by asking models to reason about contextually constrained scenarios in real world applications. As a result of his knowledge verbalization method, COPLAN is created, a considerable (counterfactual) procedure planning dataset. Using multitasking and task-specific distillation, COPLAN is subsequently used to train smaller models, PLASMA. They realize that the traditional next token prediction goal in autoregressive LMs (applied during distillation) does not give them the temporal and causal reasoning skills they need to produce high-quality plans or a way to correct their phasing errors. previous.

To overcome this difficulty, they created PLASMA+, a verifier-guided step-by-step beam search that best utilizes the multi-step plan structure. They specifically add a stepping validator into their decoding procedure to help PLASMA+ produce more semantically consistent and accurate plans over time. Through testing, they prove that their strategy brings planning skills to younger LMs. The smallest student models (of various sizes) outperform their instructor on average by 17.57% on the common planning task. Even GPT-3, a model 16 times student size, can be compared to the best student model.

Additionally, we distilled counterfactual planning skills into small-size models for the first time, achieving a 93% validity rate in human evaluation. Their model far exceeds previous work based on GPT-3 in a simulated environment in terms of runnability (17%) and accuracy (25%). When taken as a whole, its framework, consisting of the symbolic procedural distillation, decoding time algorithm, suggested tasks, and COPLAN dataset, offers significant resource and starting points for future study in planning. procedural.

review the Paper. Don’t forget to join our 22k+ ML SubReddit, discord channel, and electronic newsletter, where we share the latest AI research news, exciting AI projects, and more. If you have any questions about the article above or if we missed anything, feel free to email us at [email protected]

🚀 Check out 100 AI tools at AI Tools Club