TL;DR: In RLHF, there’s pressure between the reward studying section, which makes use of human desire within the type of comparisons, and the RL fine-tuning section, which optimizes a single, non-comparative reward. What if we carried out RL in a comparative approach?
Determine 1:
This diagram illustrates the distinction between reinforcement studying from absolute suggestions and relative suggestions. By incorporating a brand new part – pairwise coverage gradient, we will unify the reward modeling stage and RL stage, enabling direct updates based mostly on pairwise responses.
Giant Language Fashions (LLMs) have powered more and more succesful digital assistants, reminiscent of GPT-4, Claude-2, Bard and Bing Chat. These techniques can reply to complicated consumer queries, write code, and even produce poetry. The approach underlying these superb digital assistants is Reinforcement Studying with Human Suggestions (RLHF). RLHF goals to align the mannequin with human values and remove unintended behaviors, which might usually come up because of the mannequin being uncovered to a big amount of low-quality knowledge throughout its pretraining section.
Proximal Coverage Optimization (PPO), the dominant RL optimizer on this course of, has been reported to exhibit instability and implementation issues. Extra importantly, there’s a persistent discrepancy within the RLHF course of: regardless of the reward mannequin being educated utilizing comparisons between varied responses, the RL fine-tuning stage works on particular person responses with out making any comparisons. This inconsistency can exacerbate points, particularly within the difficult language era area.
Given this backdrop, an intriguing query arises: Is it attainable to design an RL algorithm that learns in a comparative method? To discover this, we introduce Pairwise Proximal Coverage Optimization (P3O), a way that harmonizes the coaching processes in each the reward studying stage and RL fine-tuning stage of RLHF, offering a passable resolution to this concern.
Background
Determine 2:
An outline of the three levels of RLHF from an OpenAI weblog publish. Be aware that the third stage falls below Reinforcement Studying with Absolute Suggestions as proven on the left facet of Determine 1.
In conventional RL settings, the reward is specified manually by the designer or supplied by a well-defined reward operate, as in Atari video games. Nevertheless, to steer a mannequin towards useful and innocent responses, defining a superb reward is just not simple. RLHF addresses this downside by studying the reward operate from human suggestions, particularly within the type of comparisons, after which making use of RL to optimize the realized reward operate.
The RLHF pipeline is split into a number of levels, detailed as follows:
Supervised Superb-Tuning Stage: The pre-trained mannequin undergoes the utmost probability loss on a top quality dataset, the place it learns to answer human queries via mimicking.
Reward Modeling Stage: The SFT mannequin is prompted with prompts (x) to provide pairs of solutions (y_1,y_2sim pi^{textual content{SFT}}(yvert x)). These generated responses kind a dataset. The response pairs are introduced to human labellers who specific a desire for one reply over the opposite, denoted as (y_w succ y_l). A comparative loss is then used to coach a reward mannequin (r_phi):
[mathcal{L}_R = mathbb{E}_{(x,y_l,y_w)simmathcal{D}}log sigmaleft(r_phi(y_w|x)-r_phi(y_l|x)right)]
RL Superb-Tuning Stage: The SFT mannequin serves because the initialization of this stage, and an RL algorithm optimizes the coverage in the direction of maximizing the reward whereas limiting the deviation from the preliminary coverage. Formally, that is performed via:
[max_{pi_theta}mathbb{E}_{xsim mathcal{D}, ysim pi_theta(cdotvert x)}left[r_phi(yvert x)-beta D_{text{KL}}(pi_theta(cdotvert x)Vert pi^{text{SFT}}(cdotvert x))right]]
An inherent problem with this method is the non-uniqueness of the reward. For example, given a reward operate (r(yvert x)), a easy shift within the reward of the immediate to (r(yvert x)+delta(x)) creates one other legitimate reward operate. These two reward capabilities end in the identical loss for any response pairs, however they differ considerably when optimized in opposition to with RL. In an excessive case, if the added noise causes the reward operate to have a wide range, an RL algorithm could be misled to extend the probability of responses with greater rewards, although these rewards is probably not significant. In different phrases, the coverage could be disrupted by the reward scale info within the immediate (x), but fails to be taught the helpful half – relative desire represented by the reward distinction. To deal with this concern, our goal is to develop an RL algorithm that’s invariant to reward translation.
Derivation of P3O
Our thought stems from the vanilla coverage gradient (VPG). VPG is a extensively adopted first-order RL optimizer, favored for its simplicity and ease of implementation. In a contextual bandit (CB) setting, the VPG is formulated as:
[nabla mathcal{L}^{text{VPG}} = mathbb{E}_{ysimpi_{theta}} r(y|x)nablalogpi_{theta}(y|x)]
By some algebraic manipulation, we will rewrite the coverage gradient in a comparative kind that entails two responses of the identical immediate. We identify it Pairwise Coverage Gradient:
[mathbb{E}_{y_1,y_2simpi_{theta}}left(r(y_1vert x)-r(y_2vert x)right)nablaleft(logfrac{pi_theta(y_1vert x)}{pi_theta(y_2vert x)}right)/2]
Not like VPG, which straight depends on absolutely the magnitude of the reward, PPG makes use of the reward distinction. This allows us to bypass the aforementioned concern of reward translation. To additional increase efficiency, we incorporate a replay buffer utilizing Significance Sampling and keep away from massive gradient updates by way of Clipping.
Significance sampling: We pattern a batch of responses from the replay buffer which include responses generated from (pi_{textual content{outdated}}) after which compute the significance sampling ratio for every response pair. The gradient is the weighted sum of the gradients computed from every response pair.
Clipping: We clip the significance sampling ratio in addition to the gradient replace to penalize excessively massive updates. This method allows the algorithm to trade-off KL divergence and reward extra effectively.
There are two alternative ways to implement the clipping approach, distinguished by both separate or joint clipping. The ensuing algorithm is known as Pairwise Proximal Coverage Optimization (P3O), with the variants being V1 or V2 respectively. You’ll find extra particulars in our unique paper.
Analysis
Determine 3:
KL-Reward frontier for TL;DR, each sequence-wise KL and reward are averaged over 200 check prompts and computed each 500 gradient steps. We discover {that a} easy linear operate matches the curve nicely. P3O has the most effective KL-Reward trade-off among the many three.
We discover two completely different open-ended textual content era duties, summarization and question-answering. In summarization, we make the most of the TL;DR dataset the place the immediate (x) is a discussion board publish from Reddit, and (y) is a corresponding abstract. For question-answering, we use Anthropic Useful and Innocent (HH), the immediate (x) is a human question from varied matters, and the coverage ought to be taught to provide an interesting and useful response (y).
We examine our algorithm P3O with a number of efficient and consultant approaches for LLM alignment. We begin with the SFT coverage educated by most probability. For RL algorithms, we think about the dominant method PPO and the newly proposed DPO. DPO straight optimizes the coverage in the direction of the closed-form resolution of the KL-constrained RL downside. Though it’s proposed as an offline alignment technique, we make it on-line with the assistance of a proxy reward operate.
Determine 4:
KL-Reward frontier for HH, every level represents a median of outcomes over 280 check prompts and calculated each 500 gradient updates. Left two figures examine P3O-V1 and PPO with various base mannequin sizes; Proper two figures examine P3O-V2 and DPO. Outcomes displaying that P3O cannot solely obtain greater reward but additionally yield higher KL management.
Deviating an excessive amount of from the reference coverage would lead the web coverage to chop corners of the reward mannequin and produce incoherent continuations, as identified by earlier works. We’re considering not solely the nicely established metric in RL literature – the reward, but additionally in how far the realized coverage deviates from the preliminary coverage, measured by KL-divergence. Due to this fact, we examine the effectiveness of every algorithm by its frontier of achieved reward and KL-divergence from the reference coverage (KL-Reward Frontier). In Determine 4 and Determine 5, we uncover that P3O has strictly dominant frontiers than PPO and DPO throughout varied mannequin sizes.
Determine 5:
Left determine shows the win charge evaluated by GPT-4. Proper determine presents the win charge based mostly on direct comparability of the proxy reward. Regardless of the excessive correlation between two figures, we discovered that the reward win charge should be adjusted based on the KL with a purpose to align with the GPT-4 win charge.
To straight assess the standard of generated responses, we additionally carry out Head-to-Head Comparisons between each pair of algorithms within the HH dataset. We use two metrics for analysis: (1) Reward, the optimized goal throughout on-line RL, (2) GPT-4, as a devoted proxy for human analysis of response helpfulness. For the latter metric, we level out that earlier research present that GPT-4 judgments correlate strongly with people, with human settlement with GPT-4 sometimes related or greater than inter-human annotator settlement.
Determine 5 presents the great pairwise comparability outcomes. The common KL-divergence and reward rating of those fashions is DPO > P3O > PPO > SFT. Though DPO marginally surpasses P3O in reward, it has a significantly greater KL-divergence, which can be detrimental to the standard of era. Because of this, DPO has a reward win charge of 49.5% in opposition to P3O, however solely 45.4% as evaluated by GPT-4. In contrast with different strategies, P3O reveals a GPT-4 win charge of 57.0% in opposition to PPO and 69.3% in opposition to SFT. This result’s per our findings from the KL-Reward frontier metric, affirming that P3O might higher align with human desire than earlier baselines.
Conclusion
On this weblog publish, we current new insights into aligning massive language fashions with human preferences by way of reinforcement studying. We proposed the Reinforcement Studying with Relative Suggestions framework, as depicted in Determine 1. Beneath this framework, we develop a novel coverage gradient algorithm – P3O. This method unifies the elemental rules of reward modeling and RL fine-tuning via comparative coaching. Our outcomes present that P3O surpasses prior strategies when it comes to the KL-Reward frontier in addition to GPT-4 win-rate.
BibTex
This weblog is predicated on our current paper and weblog. If this weblog conjures up your work, please think about citing it with:
@article{wu2023pairwise,
title={Pairwise Proximal Coverage Optimization: Harnessing Relative Suggestions for LLM Alignment},
writer={Wu, Tianhao and Zhu, Banghua and Zhang, Ruoyu and Wen, Zhaojin and Ramchandran, Kannan and Jiao, Jiantao},
journal={arXiv preprint arXiv:2310.00212},
yr={2023}
}
