🧵[7/n]

🔍 Potential Reasons

💡 We hypothesize that the in-distribution nature of training data is a key driver behind this sparsity
🧠 The model already "knows" a lot — RL just fine-tunes a small, relevant subnetwork rather than overhauling everything

🧵[6/n]

🌐 The Subnetwork Is General
🔁 Subnetworks trained with different seed, datasets, or even algorithms show nontrivial overlap
🧩 Suggests the subnetwork is a generalizable structure tied to the base model
🧠 A shared backbone seems to emerge, no matter how you train it

🧵[5/n]📊
🧪 Training the Subnetwork Reproduces Full Model

1️⃣ When trained in isolation, the sparse subnetwork recovers almost the exact same weights as the full model
2️⃣ achieves comparable (or better) end-task performance
3️⃣ 🧮 Even the training loss converges more smoothly

🧵[4/n]

📚 Each Layer Is Equally Sparse (or Dense)

📏 No specific layer or sublayer gets special treatment — all layers are updated equally sparsely.
🎯Despite the sparsity, the updates are still full-rank

🧵[3/n]

📉 Even Gradients Are Sparse in RL 📉

🧠 In PRIME, 72% of parameters never receive any gradient — ever!
↔️ Some do, but their gradients cancel out over time.
🎯 It’s not just sparse updates, even the gradients are sparse

🧵[2/n]

💡 SFT Updates Are Dense 💡
Unlike RL, Supervised Fine-Tuning (SFT) updates are much denser 🧠
📊 Sparsity is low — at most only 15.31% of parameters remain untouched.

🚨 Paper Alert: “RL Finetunes Small Subnetworks in Large Language Models”

From DeepSeek V3 Base to DeepSeek R1 Zero, a whopping 86% of parameters were NOT updated during RL training 😮😮
And this isn’t a one-off. The pattern holds across RL algorithms and models.
🧵A Deep Dive

📈 Our results:
PARC improves error detection accuracy by 6-16%, enabling more reliable step-level verification in mathematical reasoning chains.

🧵[5/n]

📊 The exciting part?
LLMs can reliably identify these critical premises - the specific prior statements that directly support each reasoning step. This creates a transparent structure showing exactly which information is necessary for each conclusion.

🧵[4/n]

🚀Our ICML 2025 paper introduces "Premise-Augmented Reasoning Chains" - a structured approach to induce explicit dependencies in reasoning chains.

By revealing the dependencies within chains, we significantly improve how LLM reasoning can be verified.

🧵[1/n]

🚩 We discovered some key gaps: Incomplete cultural coverage, issues with methodological robustness, and a lack of situated studies for real-world applicability. These gaps limit our understanding of cultural biases in LLMs. [5/7]

📚 Most studies use black-box probing methods to examine LLMs' cultural biases. However, these methods can be sensitive to prompt wording, raising concerns about robustness and generalizability. [4/7]

Moreover, following Hershcovich et al. (2022), we examined the Linguistic-Cultural Interaction in current Cultural LLM research. Notably, none of the papers we reviewed address the concept of "Aboutness." [3/7]

📊 We categorize cultural proxies into demographic proxies (such as region, language, ethnicity) and semantic proxies (such as values, norms, food habits). Current research mainly explores values and norms, leaving many cultural aspects unexplored. [2/7]

📢📢LLMs are biased towards Western Culture. Well, okay, but what do you mean by "Culture"?
In our survey of on cultural bias in LLMs, we reviewed ~90 papers. Interestingly, none of these papers define "culture" explicitly. They use “proxies”. [1/7]
[Appeared in EMNLP mains]