More Critics, Less Gain: Self-Questioning Has a Stability Limit
Counterfactual self-critique can improve compact reasoning models, but the evidence suggests critic count is a stability parameter rather than a quantity to maximize.
Cognaptus Insights
Industry thoughts, trends, and commentary.
No Click Is Not a No: Turning Passive Feedback Into Reward Signals
ImplicitRM shows how reward models can separate user preference from the likelihood of acting on it, making passive interaction logs more informative without pretending that every non-action is a negative label.
One Score, More Signals: Making Reward Models Easier to Rank and Audit
A feature-augmented reward model improves preference ranking on HH-RLHF while exposing measurable signals that product, safety, and governance teams can inspect.
Personalization Starts Before the New User Arrives
A meta-learning approach suggests that cold-start personalization may depend more on learning where user adaptation begins than on adding more user-specific model capacity.
The Reward Model Has to Move Too: R2M Tracks the Policy During RLHF
R2M shows how lightweight reward-model adaptation can track a changing policy without retraining the evaluator’s full backbone.
Buy Fewer Labels, Ask Better Questions: RLHF as an Allocation Problem
A Gemma 9B experiment suggests RLHF teams can extract substantially more value from preference budgets by adapting what they label, when they update, and which comparisons receive judgment.
One Model, Two Routes: Why Audio-Omni Unifies Audio by Splitting Its Controls
Audio-Omni suggests that consolidating multimodal audio systems works best when semantic intent and temporally precise controls remain architecturally distinct.
Same VQA Score, Different Eyes: What Fine-Grained Tests Reveal About VLMs
Broad VLM benchmarks can hide large differences in precise visual discrimination, changing how teams should evaluate models and allocate multimodal training effort.
Search With Vision, Answer With Text? What IRPAPERS Changes About Scientific RAG
IRPAPERS shows why scientific RAG systems may benefit from visual retrieval without making images the default representation for answer generation.
Think Harder, See Less? What Visual Illusions Reveal About Multimodal Reasoning
IllusionReasoning shows that extra multimodal deliberation helps explanation but can hurt simpler perception and constrained-choice tasks, making reasoning depth a routing decision rather than a default upgrade.