supervised fine-tuning

2026-10-31

RESEARCH · 1 SOURCE · arXiv cs.AI

Prospero specialized LLM outperforms base model on 250‑question ICF benchmark (arXiv:2610.38024v1)

The paper (arXiv:2610.38024v1) presents an expert‑curated 250‑question benchmark for inertial confinement fusion (ICF) covering qualitative reasoning, numerical calculation, symbolic derivation, and equation formulation, and compares out‑of‑the‑box closed‑ and open‑source LLMs to Prospero, a model specialized via domain‑adaptive pretraining and supervised fine‑tuning. Prospero improved its aggregate score by 10.7 percentage points over its base model; frontier models scored highest overall but still made critical physics and numerical errors, retrieval helped a smaller model but not frontier models, and the authors warn that benchmark gains do not establish scientific reliability.

6.0

2026-10-31

RESEARCH · 1 SOURCE · arXiv cs.AI

LLMON post-training teaches explicit instruction/data separation and cuts prompt-injection success

The paper introduces LLMON (LLM Object Notation), a post-training supervision format that enforces explicit identifier-based execution by separating instructions from data. Using a 4.6B-token structured corpus (including RAG and tool-calling data with distractor-infused variants), the authors fine-tune Granite-4.1 3B and 8B models via supervised fine-tuning followed by reinforcement learning and report that LLMON supervision reduces indirect prompt-injection attack success rates from 53–100% down to 0.02–5.6% while preserving general capabilities; they also propose attention-based interpretability metrics that reveal how attention patterns change under structured inputs.

8.0

2026-09-30

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Systematic study finds effectiveness–fluency trade-off in LLM conditioning

The paper systematically evaluates a range of LLM conditioning methods for both concept injection and removal, showing that many efficient steering techniques significantly harm generation fluency. It also reports that activation-steering methods work much less well on instruction-tuned models than on base models, while prompting and supervised fine-tuning are viable for injection (but weaker for removal), and that inexpensive textual metrics correlate well with costly LLM-as-judge scores. Published in TMLR (Sep 18, 2026).

7.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

Forecast-Dojo: replayable benchmark and training environment for LLM forecasting agents (arXiv:2609.28876v1)

Forecast-Dojo is a replayable environment that pairs 1,568 resolved Polymarket prediction questions with 18.8M dated news articles so agents can research events and revisit forecasts at successive historical dates. The paper evaluates 12 models, finds research tools reduce Brier score and that forecasts improve as events unfold (though all models still trail historical market forecasts), and provides interaction trajectories plus a supervised fine-tuning proof of concept.

6.0

2026-08-26

RESEARCH · 1 SOURCE · Apple Machine Learning Research

PROOF-Gen: From optimized training data to improved distillation for tool-calling

The paper presents PROOF-Gen, a method that focuses on generating optimized training data to improve supervised fine-tuning of student models distilled from teacher-generated trajectories, aiming to overcome limitations of the common generate-and-filter pipeline that leaves hard failure cases unaddressed. The authors report that on τ 2-bench, 57% of teacher trials fail and roughly two-thirds of those failures are near-misses (most tool calls correct but subsequently undone), motivating the need to extract signal from failures rather than discarding them.

7.0