When Does LLM Reasoning Actually Help? — Detectable via Entropy Phase Transitions

Chain-of-Thought (CoT) reasoning is widely used as a standard technique for eliciting LLM capabilities, but in practice it often provides little benefit on fact-checking and open-ended tasks, and in some cases simply increases token consumption. New research addresses this paradox by answering the question: when does reasoning actually help?

The paper "When Do LLMs Reason?" reframes LLM reasoning not as a static property of tasks or models, but as a **dynamic decoding state during generation**. Analyzing the dynamics of entropy in the early stages of generation, the authors found that tasks where CoT is effective exhibit a consistent decrease in entropy, while other tasks show instability or a rising trend.

This behavior can be interpreted as a **phase-transition-like phenomenon**. In other words, whether an LLM enters "reasoning mode" may be determined at a very early stage of generation. This finding holds promise as a practical signal for allocating reasoning resources appropriately.

Model Collapse Explained Linguistically Through "Cultural Evolution"

The "model collapse" phenomenon, in which LLMs retrained on their own outputs gradually degrade, had been characterized statistically, but it remained unclear which linguistic structures collapse, in what order, and why.

"Model Collapse as Cultural Evolution" shows that **iterated learning theory**, well known in the field of cultural evolution, fills this gap. The authors ran experiments in which LLaMA-2-7B and Mistral-7B self-trained over 10 generations in English, German, and Turkish.

The most important finding is that **compositionality follows a non-monotonic trajectory**: with unfiltered self-training, compositionality first rises and then declines. This signature appears even with fully regular initial data and cannot be explained by denoising. It suggests that model collapse is not mere statistical degradation but a systematic reorganization of linguistic structure.

Do VLM Benchmarks Really Test "Vision"?

The premise that Vision-Language Model (VLM) benchmark scores reflect visual understanding has now been called into serious question.

The team behind "Seeing without Looking" started from a striking observation: removing the vast majority of image tokens causes only a minor drop in performance on widely used hallucination benchmarks. They then probed open-source VLMs at multiple granularities, including global visual degradation, local occlusion, question reformulation, expanded answer spaces, and decision-level analyses beyond accuracy.

They further analyzed the geometric structure of visual tokens layer by layer. The results show that many VLMs **rely heavily on textual cues and prior knowledge**, with far less dependence on actual visual evidence than appearances suggest — a finding that forces a fundamental rethink of how VLMs are evaluated.

LLMs Communicating Without Text — Latent Cache Flow

Communication between LLM agents today happens through text, which incurs substantial latency and information loss. Recent Cache-to-Cache (C2C) research tries to solve this by transforming KV caches, but its adapters are large and it requires identical contexts.

"Latent Cache Flow (LCF)" is a new method that tackles this problem. By **jointly transforming and compressing** keys and values, it shrinks the adapter to roughly 4% the size of C2C. Moreover, to work even between models with different contexts, it is designed to send summaries of the new information the target model lacks.

In initial experiments, an LCF adapter of just 13MB outperformed text communication, suggesting it could dramatically improve communication efficiency in multi-agent systems.

AI Has a "Deterministic Horizon" — Proving the Limits of Reasoning

As LLMs write software, draft legal documents, and generate clinical records, it is essential to consider the fundamental limits of computation.

"The Deterministic Horizon" is a paper that turns impossibility results — Turing's limits, Arrow's impossibility theorem, and the No Free Lunch theorem — from theoretical curiosities into **practical design rules**. Its main contribution is a proof of an **accuracy ceiling determined by architecture alone**.

Once reasoning depth passes a critical point, no amount of learning, no adapter rank, and no loss function can move this ceiling. This "Deterministic Horizon" can be computed from the number of layers and the embedding width before deployment, and was measured between 19 and 31 across 12 Transformer architectures. Even fine-tuning on optimally long traces yields improvements of less than 4 points, the paper reports.

This marks an important step toward predicting AI capability limits in advance and building them into design.

---