Hadal Research
§ Papers

Publications

Every entry carries its records and the analysis code that produced its numbers. Preprints are labelled as preprints.

15 NOVEMBER 2026
v1

Framing Newlines: how tool-call parsers rewrite the files agents write

Kyssta · PREPRINT · NOT PEER REVIEWED

Every open-weights agent model that emits tool calls in the Qwen/LFM3 XML framing writes each parameter value between one newline on either side. We rendered 1,762 real files as write_file calls and parsed them back with five parser implementations found in shipped code. Parsed by frame the format is invertible: 1,759 of 1,762 files return byte-identical, and the three failures contain the unescaped terminator. Parsed with .strip(), 15 survive. On a benchmark that grades file contents, a perfect model therefore scores 0.9% under the dominant client-side parser and 99.8% under the rule the reference server implementation uses — a 99-point swing with no model in the loop.

Read on the blogPaper (PDF)Code & records

27 OCTOBER 2026
v6

Changing One Word: a paired test of anchor direction in list indexing

Kyssta · PREPRINT · NOT PEER REVIEWED

For an odd-length list, the middle item can be requested two ways that name the same item at the same ordinal: “5th position from the beginning” and “5th position from the end”. Only the anchor word differs. On Qwen2.5-1.5B-Instruct with 9-item lists, the backward wording reduces accuracy by 33.3 percentage points (paired difference −33.3 pp, 95% CI [−51.2, −11.6], exact McNemar p = 0.00781), and the estimated effect differs between the two question templates tested. The reduction is established; the pre-specified 15-point practical margin is not cleared.

Read on the blogPaper (PDF)Code & records