LLMs

Researchers introduce Iris-mini and Iris-pro, search agents scaled to 35B and 397B parameters, trained using a novel data pipeline. They construct multi-hop questions by rewriting entities into descriptive references to prevent string-matching shortcuts, ensuring models rely on retrieved evidence. The policy is further refined via reinforcement learning against live search environments to handle complex, multi-step information retrieval tasks.

• Uses entity graph distillation to generate complex multi-hop questions from web hyperlinks.
• Rewrites entities to descriptive references, forcing models to use evidence rather than string matching.
• Applies reinforcement learning against live search to optimize policy performance on real-world queries.
• Filters training trajectories at both turn and sequence levels to ensure high-quality SFT data.

This research addresses unfaithful LLM explanations by identifying two failure modes: incompleteness, where relevant factors are omitted, and unsoundness, where irrelevant factors are cited. The proposed method operates at test time without requiring retraining, using a removal-based approach to evaluate which input features actually drive the model's answer and explanation. This allows for auditing and improving the alignment between stated reasoning and actual model behavior.

• Distinguishes between incomplete and unsound explanations for precise auditing.
• Improves faithfulness at test time, avoiding costly model retraining.
• Uses feature removal to isolate causal factors in LLM decision-making.
• Provides a concrete method to verify if explanations reflect true reasoning.

Researchers challenge the assumption that effective post-training requires massive token consumption. When using on-policy distillation with Qwen3, they found that reasoning skills can be incentivized by supervising only one or two tokens per trajectory. This sparse approach, representing just 0.05% of total tokens, matches or exceeds the performance of full-token training methods.

• Post-training does not need to be token-intensive to be effective.
• On-policy distillation allows for extremely sparse supervision signals.
• Supervising 1-2 tokens per trajectory yields strong reasoning gains.
• Sparse methods achieve performance equal to or better than full-token training.
• Qwen3 family results suggest broad applicability of this phenomenon.

A new distributional theory explains how LLMs incorporate external evidence by adjusting the probability distribution of initial answers rather than discarding them. The model relies on a receiver prior weight and candidate evidence tilt to predict persuasion and error integration. Key findings include that models favor their own characteristic errors and that identical evidence can improve weaker models while harming stronger ones.

• Evidence shifts answer distributions via prior weight and evidence tilt parameters
• Models are more persuaded by candidates already probable to their initial state
• LLMs integrate characteristic errors from their own training more readily than foreign errors
• Identical evidence can improve weaker models but degrade performance in stronger ones
• Theory validated across ten million data points for robust predictive power

A new study on Qwen3.5 and Falcon-H1 models uses cache-level interventions to isolate the functions of attention and recurrent states. Results show attention is strictly required for retrieving precise facts, while the recurrent state governs output language and persona. Swapping these components confirms that factual accuracy vanishes without attention, whereas stylistic traits persist even when KV caches are removed.

• Attention is the sole channel for exact factual retrieval in hybrid models
• Recurrence preserves output style and persona, independent of KV cache
• Split-prefill and state-swap tests prove functional separation of channels
• Hybrid architectures leverage attention for memory and recurrence for voice

Merging LoRA adapters into native 4-bit microscaled models (NVFP4/MXFP4) is typically lossy because re-quantization resets weights to the nearest discrete code plane. This approach introduces Scale-QLoRA, a method that preserves adaptation quality during the merge by maintaining code invariance. Without this technique, naive merging can degrade performance by up to 39 percentage points.

• Native 4-bit merging destroys LoRA gains unless code-invariant techniques are used.
• Re-quantization snaps weights to the nearest discrete E2M1 code plane, losing precision.
• Scale-QLoRA enables stable adapter merging without performance degradation.
• Critical for deployment pipelines involving NVFP4 or MXFP4 microscaling formats.

A Hacker News thread argues that relying on large language models for authoring posts reveals a lack of original thought or deep understanding. The discussion frames AI-assisted writing not as a productivity tool, but as a signal of intellectual negligence that undermines professional credibility. Participants emphasize that technical audiences value authentic, self-authored insights over polished, AI-generated content.

• AI-generated content may signal shallow expertise rather than efficiency
• Community perception values original thought over polished output
• Relying on LLMs for authorship can damage professional credibility
• Technical audiences prefer authentic, self-authored technical posts