Hello OpenAI Community,
I have recently completed two research papers exploring explicit knowledge boundaries and calibrated selective prediction to eliminate model hallucinations..
I am ready to submit them to arXiv but am currently blocked by the first-time author endorsement requirement for the cs.AI (Artificial Intelligence) section.
Abstract Summaries:
-
**Paper 1:**RESEARCH_PAPER [Large Language Models (LLMs) exhibit a consistent behavioral “nature”: they love to chat.
They answer every prompt, at length, with confidence — even when they do not know the
answer. We argue this is not a bug but the inevitable consequence of the goal their
training embodies: to be able to do everything. A system optimized to always produce
the best next token, on data where a response is always present and always confident,
develops a nature in which silence, refusal, and honest ignorance do not exist.
Hallucination is the shadow cast by this nature.This paper proposes an alternative design goal — freedom through truth — and a
concrete framework for achieving it. Instead of maximizing capability, we train a model
whose knowledge is a bounded, fully-known inheritance (a curated knowledge base),
whose speech is a choice rather than a compulsion (explicit speak / verify / stop
mechanics), and whose honesty is native to its training data rather than bolted on
with post-hoc guardrails. We introduce three artifacts: (1) Heartly, an early
prototype experiment with decide-to-speak and verify-before-claim mechanics, whose
instructive failure mode motivates this framework; (2) the Nature Profile, a
declarative specification that translates chosen principles into dataset generation
rules and tokenizer schemes; and (3) the Knowledge-Base Dataset Organizer, a tool
that maintains deduplicated, source-tracked knowledge and renders it — through a
Nature Profile — into training datasets of any format. We show why honest abstention is
only trainable when the knowledge boundary is explicit, and why a bounded model can be
more trustworthy, more restful, and — in a meaningful sense — more free than an
unbounded one.] -
**Paper 2:**RESEARCH_BRIEF_TECHNICAL [Hallucination under distribution shift is usually attacked at inference time (RAG, calibration probes, refusal RLHF). We propose attacking it at dataset construction time, and we identify the specific structural reason current SFT data cannot teach abstention: the knowledge boundary of the training corpus is implicit, so correct abstention examples cannot be generated systematically.
-
If the fine-tuning corpus is derived from an explicit, enumerable knowledge base (KB), the complement of the KB becomes queryable — and abstention examples can be generated exactly at the knowledge boundary, at any desired ratio, with gold labels that are correct by construction rather than by annotation.
-
We propose (1) a falsifiable hypothesis, (2) a cheap experiment (single-GPU-days scale) with clear metrics and baselines, (3) a small negative result from a prior attempt that sharpens the hypothesis, and (4) an open tooling effort (KB → dataset compiler with declarative behavioral specifications).
-
The interesting question is not “can a model say I don’t know” — R-Tuning showed it can. The question is whether boundary-exact, systematically generated abstention data produces calibrated selective prediction that generalizes along the boundary, rather than a surface refusal style.]
If there are any active arXiv contributors in this forum who would be willing to review my abstracts and endorse my account, I would deeply appreciate your support.
Thank you for your time!