FATE Series (SABER & CASTER) Debuts at ACL 2026: Advanced LLM Reasoning

SABER: Switchable and Balanced Training for Efficient LLM Reasoning

Existing reasoning‑enhanced LLMs (e.g., Chain‑of‑Thought, Long‑Chain Reasoning) often exhibit overthinking: even trivial queries such as “1+1=?” generate unnecessarily long reasoning chains, increasing latency and cost.

SABER introduces a token‑budget‑constrained, switchable training paradigm. It uses GRPO reinforcement learning to directly optimise reasoning without a pre‑training SFT warm‑up, enabling a single model to support four discrete reasoning modes.

Experiments on MATH, GSM8K and MBPP show that FastThink reduces reasoning length by 65 %–80 % while preserving or improving accuracy; NoThink incurs minimal performance degradation compared with the base model.

Paper: https://arxiv.org/abs/2508.10026

Code: https://github.com/bilibili/saber_rl

CASTER: Community‑Aware Assessment of Social Textual Engagement and Resonance

Traditional video‑quality‑assessment (VQA) metrics focus on pixel‑level fidelity, whereas UGC quality is driven by community consensus. CASTER takes multimodal video information (cover, key frames, title, tags, ASR, etc.) and simulates reactions from diverse audience personas to predict community acceptance.

Social‑CoT reasoning mechanism

Instantiate multiple audience personas (e.g., veteran enthusiast, casual passer‑by, newcomer, critical veteran).

For each persona, simulate emotional response paths: identify impactful segment, generate a plausible comment, etc.

Aggregate simulated responses using a Skellam‑scoring consensus to decide whether the content will resonate positively.

MEDEA framework

Stage 1: Extract real community wisdom from Bilibili using a teacher model (Gemini) to generate 54 K annotated Social‑CoT reasoning paths.

Stage 2: Supervised fine‑tuning (SFT) aligns visual cues and textual signals with social interpretations.

Stage 3: Reinforcement‑learning alignment (GRPO) with four composite rewards: format reward, label reward, cognitive‑diversity constraint, and social‑alignment reward (matching high‑up‑vote comments).

CASTER‑Bench dataset

1 485 UGC videos covering 30 categories (life, knowledge, gaming, food, tech, dance, etc.).

Average length 442 seconds (total 182.5 h), far longer than existing VQA clips (8‑10 s).

Each entry includes full multimodal metadata (video, cover, title, tags, ASR, etc.).

Experimental results

On CASTER‑Bench, MEDEA achieves F1 = 0.650, precision = 0.603, recall = 0.705, surpassing the strongest baseline (GPT‑5.2 reasoning, F1 = 0.555) by +17.1 %.

Failure‑mode analysis:

Traditional VQA methods (FastVQA, DOVER, MaxVQA) obtain low F1 (0.33‑0.41) because they assess visual quality rather than community relevance.

Standard LLMs (GPT‑5.2, Claude‑4.5‑Opus) show high recall (>90 %) but low precision (~30 %) due to “generosity bias”: they can find positive aspects in any video but lack discriminative social judgment.

Reasoning‑enhanced LLMs improve slightly (max F1 = 0.555) but logical reasoning does not equate to social cognition.

Direct Social‑CoT prompting without fine‑tuning yields F1 = 0.508, indicating that the reasoning pattern helps but specialised training is required to internalise community standards.

CASTER has been integrated into Bilibili’s content pipeline, allowing early identification of videos with high community resonance before comments appear.

Paper: https://arxiv.org/abs/2606.01897

Code: https://github.com/bilibili/medea_rl

Model: https://huggingface.co/IndexTeam/MEDEA

Dataset: https://huggingface.co/datasets/IndexTeam/CASTER-Bench