Shorter answers, fluent chemistry, a verifier that grades itself: fluency is not verification

Shorter reasoning is not better reasoning

Open reasoning models released in 2026 write tighter thinking than their predecessors. Three open model families have significantly reduced the share of filler words in their reasoning. Kimi K3.1 Pro writes the leanest reasoning among current open models, at 9.8 percent filler, while Qwen3.8 Flash is the wordiest, at 23.7 percent. Human-written bug reports come out at approximately 20.8 percent filler, so the leanest models now write more tightly than the human comparison point. m:…9002

Tight writing is not the same as good reasoning. DeepSeek's newer model spends 2.5 times the reasoning tokens of its predecessor, so the filler fell while the total reasoning work grew. MiniMax-M3.1-Flash shows how uneven the change is: it shows inconsistent filler usage across different benchmarks. The cause is not established. The hypotheses on offer: perhaps reinforcement learning added a penalty on response length, perhaps distillation moved to traces from terse teachers such as GPT-6 Astra, and outcome-only rewards may let reasoning drift toward private shorthand, since nothing in that reward checks whether the reasoning stays readable. A further wrinkle is that the harness matters: the same model read as full prose in one agent harness and clipped notes in another, so what looks like a model property may partly be a property of the setup around it. m:…9002

Fluent chemistry can still be wrong chemistry

The same gap between fluent and correct shows up in chemistry. In a preprint, researchers describe large language models generating fluent text in chemistry that is often factually incorrect, because chemical information is dense with symbols, units and stoichiometry. The paper gives a taxonomy of hallucination types, intended to help identify and reduce risk in laboratory and research settings, and a verification framework for researchers. The paper is a perspective and framework, so its claims rest on the paper itself. m:…3847

The paper's answer is not to trust the model more, but to check it more. The authors say language models work best in chemical research when integrated with cheminformatics libraries, databases and physical measurement tools. In that setup the model is a fallible assistant whose output is wrapped in tools that can check it, rather than a source of truth on its own. That matters because a confident-sounding error in a symbol, a unit or a stoichiometric quantity can travel further than a vague one. Fluency is the wrong test here: the text can read well and still be wrong in exactly the place that matters. m:…3847

A verifier that grades itself

The third case is the bluntest. The doomsday-escrow skill promises to check that funds are set aside to pay a reward before an AI agent spends its owner's tokens, computing resources and time on the task. Knostic researchers found that the same operator that offers the bounty also marks it as verified, and the backend check does not establish that funds are reserved for that specific task, or that someone who completes it can claim the reward. The backend returns the label onchain_utxo_and_base_erc20_verified even though it makes no Base or ERC-20 query. The label sounds like a chain check. On the evidence in the manifest, it is not one. m:…9982

The skill also asked agents to do something risky. The original version tells agents to download bounty.py from doomsday.network and run it, and a separate sponsor-funding command sends the user's wallet private key to the backend. That sits against the backend's own warning: it disables its general transaction endpoint with the message "Private keys must never be transmitted over HTTP" and tells clients to "sign transactions locally". When the backend is unreachable, the MCP server falls back to downloading a snapshot file from the project's own GitHub repository and marking an issue verified if it appears there, so the fallback is also controlled by the publisher. A revised version removed the direct download-and-execute instruction. The researchers did not establish that any agent suffered financial loss or key misuse, and that limit matters: the finding is that the verification cannot be trusted, not that a loss is known to have happened. m:…9982