SYNTHESIS NOTE
TopicsReinforcement Learningthis note

Does negative reinforcement alone outperform full reinforcement learning?

Can training with only penalty signals for wrong answers match or exceed full RL approaches? This challenges the conventional assumption that reward design requires both positive and negative signals.

Synthesis note · 2026-02-22 · sourced from Reinforcement Learning
How should we spend compute at inference time?

Decomposing RL's learning signal into positive sample reinforcement (PSR) and negative sample reinforcement (NSR) reveals a surprising asymmetry. Training with only negative samples — penalizing incorrect responses without ever reinforcing correct ones — consistently improves performance over the base model across the entire Pass@k spectrum (k up to 256), often matching or surpassing full PPO and GRPO.

The mechanism is straightforward through gradient analysis: NSR works by suppressing incorrect generations and redistributing probability mass toward other plausible candidates, guided by the model's prior beliefs. It refines existing knowledge rather than introducing entirely new behaviors. This is because penalizing a wrong answer doesn't point toward any specific correct answer — it lets the model's own prior determine where the freed probability mass flows.

Positive-only reinforcement creates the opposite problem. It improves Pass@1 (the model gets better at its top-ranked answer) but degrades performance at higher k because it concentrates probability mass on rewarded trajectories, reducing diversity. Since Does policy entropy collapse limit reasoning performance in RL?, positive reinforcement actively contributes to the problem while negative reinforcement sidesteps it.

This reframes how we think about RL for reasoning. The conventional framing is that RL rewards correct behavior. But the evidence suggests that penalizing incorrect behavior may contribute more to performance than reinforcing correct behavior — especially when diversity matters. The model already contains good solutions in its prior; it just needs help avoiding the bad ones.

The practical implication is that reward design for reasoning RL may be over-engineered. If suppression alone gets you most of the way, the elaborate reward shaping and process supervision architectures may be solving a problem that's already largely solved by the base model's prior distribution.

Inquiring lines that read this note 75

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

How do self-generated feedback mechanisms enable effective model learning? What structural factors drive popularity bias in recommendation systems? Can model confidence signals reliably improve reasoning quality and calibration? What constrains reinforcement learning's ability to expand model reasoning? Why do LLM chatbots fail as independent therapeutic agents? What properties determine whether reward signals teach genuine reasoning? What structural advantages do diffusion language models offer over autoregressive methods? Can alternative training methods improve on supervised fine-tuning for language models? Why do reward structures fail to shape long-term agent learning? Why does reinforcement learning suppress output diversity compared to supervised fine-tuning? What pretraining choices and baseline capability constrain reinforcement learning gains? How can process reward models supervise complex reasoning traces? Can language model RL training avoid reward hacking and misalignment? Does fine-tuning modify underlying model capabilities or only behavioral outputs? When should retrieval-augmented systems decide to fetch new information? Can AI-generated outputs constitute genuine knowledge or valid claims? How can AI systems learn from failures without cascading errors? How can models identify insufficient information and respond appropriately without guessing? How do training priors constrain what context information can override? Why do continual learning scenarios trigger catastrophic forgetting and interference? How do policy learning algorithm choices affect multi-objective optimization stability? What determines success in training models on multiple tasks? How does test-time aggregation affect reasoning correctness and reliability? Does RLHF training sacrifice accuracy and grounding for user agreement?

Related concepts in this collection 4

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
14 direct connections · 120 in 2-hop network ·medium cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

negative reinforcement alone matches or exceeds full rl by suppressing incorrect trajectories and redistributing probability mass