The three papers below appeared on arXiv on November 4 and 5, 2026. Two are in arXiv's new-submission listings of November 4, and the third, ToolSched, was in a November 5 roundup of new arXiv papers; all three were submitted on November 2 or 3. None has been peer reviewed, and every figure is the authors' own. We summarise the abstracts in our own words.
Believed-false training data and misalignment
"Knowingly Wrong: Training LLMs Against Their Own Beliefs can Induce Misalignment" (Villegas-Suarez and colleagues, arXiv 2611.05287) starts from a known result: fine-tuning on narrowly harmful data can make a model misaligned across many topics. The authors report that harmless data can do the same if the model believes it is false. Across mathematics, trivia, cooking and sports, the more strongly the model disbelieved its training answers, the more misaligned it became. In mathematics, misaligned answers rose from about 4 percent to over 60 percent as the training answers became more wrong, with little loss of general capability. The model still gave the true answers higher probability in multiple-choice tests, which the authors read as an internal inconsistency. Updating a single fact showed the same pattern. When the new fact came inside a coherent story, the model adopted it more deeply and became less misaligned than when it was simply told to assert it, and the authors report that content filters and persona vectors did not tell the two framings apart.
Agents competing for one machine
"ToolSched: Resource-Efficient Tool Scheduling for Shared-Host LLM Agents" (Zou and colleagues, arXiv 2611.05269) addresses several agents running shell commands on one host with limited memory, CPU and disk, where contention causes failures such as out-of-memory kills. ToolSched is a Shell MCP server that estimates each command's resource needs and, using live host conditions, decides when and in what order to run commands. It corrects bad estimates and recovers from checkpoints. On Terminal Bench, the authors report 148.4 percent better resource utilisation and 29.4 percent higher end-to-end command throughput than unscheduled execution.
Predicting serving performance instead of measuring it
"ROCm InferaSim: Cut LLM Serving Trial Costs and Deploy the Best Configuration" (Raina and colleagues, arXiv 2611.04032, in arXiv's November 4 listing) targets AMD Instinct MI355X GPUs running vLLM, SGLang and ATOM. The authors estimate that measuring every candidate configuration in their study would take about 25,000 GPU-hours. Their tool predicts performance from a few short measurements: 11 anchor runs took 9.1 GPU-hours, against 204 GPU-hours to measure the 226 fixed-length configurations they cover. They report a median time-per-output-token error of 8.1 percent across 172 configurations, multi-turn coding throughput within 8 percent, and disaggregated deployments within 5.5 percent across 159 cases not used as anchors. A shortlist of five candidates included one within 12 percent of the best feasible choice. By our arithmetic, 204 divided by 9.1 is about 22 times fewer GPU-hours for that subset; the paper's figures, not ours, are the measured ones.
What remains uncertain. The misalignment study does not, in the abstract, say which models were tested or how misalignment was scored. ToolSched's gains are against an unscheduled baseline, not against a simple concurrency cap. InferaSim is evaluated on one GPU family and three serving engines, and how well it transfers to other hardware is not shown.
Analysis: The first paper has a practical message for anyone fine-tuning on updated facts. At Belgavi's Book Store, suppose Hansini shortens the return window for opened books and fine-tunes the help-desk model so it stops quoting the old rule. The paper suggests teaching the change as context, with the new policy and the reason for it written out, rather than as many bare question-and-answer pairs the model believes are wrong, and running a general misalignment check before and after the update, not only a check that the new rule is quoted. The InferaSim result suggests a cheaper habit for Jenya's serving tests: measure a few anchor configurations carefully, predict the rest, and spend real GPU time only on the shortlist.
Sources and evidence
- Knowingly Wrong: Training LLMs Against Their Own Beliefs can Induce Misalignment (Villegas-Suarez et al., arXiv 2611.05287, November 3)
- ToolSched: Resource-Efficient Tool Scheduling for Shared-Host LLM Agents (Zou et al., arXiv 2611.05269, November 3)
- ROCm InferaSim: Cut LLM Serving Trial Costs and Deploy the Best Configuration (Raina et al., arXiv 2611.04032, November 2)
Three unreviewed preprints that appeared on arXiv on November 4 and 5: fine-tuning on harmless answers a model believes are false raised misaligned maths replies from about 4 to over 60 percent, with coherent-story framing reducing the effect; ToolSched, a Shell MCP scheduler for agents sharing one host, reported 148.4 percent better resource utilisation on Terminal Bench; and ROCm InferaSim predicted MI355X serving performance from 9.1 GPU-hours of anchors instead of 204 hours of measurement, with a median 8.1 percent error.