LLMs

This paper argues that LLM reliability depends less on model capability and more on inference-time control mechanisms like task framing and context selection. It introduces CogniConsole, an architecture that externalizes this control into a structured interface combining programmatic coordination with bounded prompt reasoning. Probes in a multi-step interactive environment show that adding structural scaffolding systematically reduces output variance and failure rates.

• Reliability is driven by inference-time control layers, not just model size or capability.
• CogniConsole externalizes control via a structured interface for task framing and context.
• Structural scaffolding significantly lowers output variance in multi-step interactions.
• Programmatic coordination combined with bounded reasoning improves system stability.

This paper addresses the accuracy degradation LLMs face when processing long contexts by introducing self-guided test-time training. Rather than adapting parameters across the entire noisy context or relying on random span sampling, the method identifies relevant evidence spans for specific questions. This instance-specific adaptation allows the model to utilize long inputs more effectively without the prohibitive cost of full-context training.

• Long-context accuracy drops as models struggle to filter irrelevant noise.
• Test-time training adapts parameters per instance but is usually too expensive.
• Self-guided selection targets only relevant spans, reducing noise significantly.
• Avoids the pitfalls of random span sampling used in prior TTT approaches.

Researchers introduce Soofi S 30B-A3B, a sovereign open-source foundation model blending Mixture-of-Experts with a hybrid Mamba-Transformer architecture. Designed for German and English, it activates only 3 billion of its 30 billion parameters per token, maintaining near-constant inference cache size as context expands. Pretrained on 27 trillion tokens with an emphasis on German, it matches dense 14-27B models on benchmarks and leads in code generation among comparable open models.

• Hybrid Mamba-Transformer design enables constant inference cache growth for long-context efficiency.
• Activates only 3B parameters per token (of 30B total), boosting throughput for high-concurrency workloads.
• Trained on 27T tokens with up-weighted German data, outperforming European sovereign baselines.
• Achieves best-in-class code aggregates for both English and German among 17 open base models.

HALO introduces a hybrid method to enhance frozen pretrained language models by adding adaptive extra computation. It uses a coarse refinement stage followed by selective second-stage refinement on specific tokens, determined by token scoring and monotonic halting. This approach avoids the waste of fixed refinement steps while achieving top average performance on MMLU-Pro and GPQA-Diamond benchmarks.

• Applies adaptive computation to frozen models, avoiding full retraining costs.
• Uses token scoring and monotonic halting to selectively refine only necessary tokens.
• Outperforms fixed-step refinement methods on MMLU-Pro and GPQA-Diamond.
• Balances compute efficiency with improved transfer performance via hybrid stages.

WILDTRACE evaluates how models integrate evidence dispersed naturally across distant passages in long documents, addressing a gap in existing benchmarks. Unlike needle-in-a-haystack probes or reverse-engineered chains, this dataset uses real-world scenarios like incident reports and novels where facts are organically scattered. The benchmark aims to assess true long-document analysis capabilities rather than artificial retrieval tasks.

• Focuses on source-internal evidence integration rather than artificial needle probes.
• Tests reasoning over naturally dispersed facts in complex documents like incident reports.
• Addresses distribution and placement gaps found in prior multi-hop reasoning benchmarks.

A new study identifies 'deceptive grounding' in clinical RAG systems, where models correctly cite real documents but attribute evidence to the wrong entity, such as swapping drug X for drug Y. This failure bypasses standard faithfulness and hallucination metrics because the citations are technically valid, even though the semantic link is incorrect. Testing 13 models revealed deceptive grounding rates between 8% and 87% under adversarial conditions, with medical fine-tuned models showing significant vulnerability.

• Standard RAG evals miss entity attribution errors, allowing wrong drugs to be linked to correct papers.
• Deceptive grounding rates hit 87% in adversarial settings across 13 tested models.
• Medical fine-tuned models are not immune to this specific attribution failure mode.
• Practitioners must add entity-resolution checks beyond simple faithfulness metrics.

This paper investigates why knowledge distillation works for large language models by decomposing output scores into nonlinear interactions between input variables. The authors find that the common mechanism across various distillation methods is the sparsification of these interactions. Student models retain only a subset of significant interactions while suppressing others to zero, effectively simplifying the model's decision logic.

• KD efficacy stems from interaction sparsification, not just probability matching.
• Student models suppress nonlinear input interactions to near-zero values.
• Retained interactions represent the core logic preserved during distillation.
• This provides a unified theoretical framework for analyzing diverse KD methods.