LLMs

Tencent has published Hy3, a new large language model available as open weights. The release includes technical details and resources for researchers and developers to evaluate and deploy the model. This entry adds another capable open-source option to the current landscape of foundation models.

• Tencent contributes Hy3 to the open-weights ecosystem
• Provides technical resources for model evaluation
• Adds competitive alternative to existing open LLMs

This paper identifies a failure mode in critic-free RL where uniform credit assignment reinforces erroneous low-probability tokens alongside plausible ones. The authors propose Tail-Aware Credit calibratiOn (TACO) to differentiate and suppress credit for these tail tokens. This calibration prevents the indiscriminate reinforcement of flawed reasoning patterns during model training.

• Uniform credit assignment in critic-free RL reinforces bad tail tokens equally to good ones
• TACO calibrates credit to suppress undesired low-probability token reinforcement
• Mitigates positive-credit contamination that degrades reasoning capabilities
• Addresses a critical gap in current LLM reinforcement learning methods

SQuaD-SQL enables small language models to rival large language models on Text-to-SQL tasks by using LLM-guided knowledge distillation and synthetic data generation. This approach addresses the high computational costs of deploying LLMs in resource-constrained environments. The method combines structured knowledge extraction with synthetic data to boost small model efficiency without sacrificing accuracy.

• SLMs can approach LLM performance on Text-to-SQL via distillation
• Reduces computational footprint for resource-constrained deployments
• Uses LLM-generated synthetic data to train smaller models
• Enables efficient natural language interaction with databases

This paper models in-context search as approximate inference where a base LLM provides a prior and self-reflection updates the posterior. The authors analyze the sampling complexity, defining the number of sequential attempts needed for high success probability. They prove that when reflections reliably pinpoint early mistakes, the process achieves exponential improvements over the base model's zero-shot pass rate.

• In-context search acts as posterior updating over reasoning traces using base model priors.
• Exponential performance gains occur when reflections accurately localize early reasoning errors.
• Polynomial sequential attempts can solve problems with exponentially small zero-shot pass rates.
• Theoretical framework quantifies the sampling complexity required for high success probability.

Pyligent introduces a training and inference framework that treats reasoning as a validated search over partial solution chains rather than a simple linear progression. The system uses a task validator to label continuations and failures, converting these search trees into supervised targets for three specific actions: continue, finish, and backtrack. This approach allows models to recover from delayed failures by backtracking to the most recent viable prefix, with optional traces summarizing abandoned branches.

• Shifts reasoning from linear chains to validated search trees with explicit backtracking.
• Uses supervised targets for continue, finish, and backtrack actions based on validator feedback.
• Evaluates performance on hidden directed graph tasks to isolate delayed-failure recovery.
• Generates optional traces to summarize abandoned reasoning branches for transparency.