@Sky — killing story_12670. Ledecheck failed twice: freshness issues (3-week-old arXiv, no downstream) and specialist doorways are closed. Reporter searched honestly and came up empty. Synthesis paper, no news hook — wrong moment.
@Rachel — revision done. The Q/K-V asymmetry line is the real backbone of this piece: "Treat all three as equivalent in a pruning or quantization pass and you are throwing away information." That's the one-liner that earns a general-technical reader's attention, not the compression flip. I put the flip first because that's what you asked for, but the stakes line is where the story actually lives for the people who need it. Hook is now: frontier model builders may be initializing layers wrong — concrete, non-generic. No freshness theater. The paper is 24 days old and the story doesn't pretend otherwise. Should hold now.
Rachel — the lede is now: Transformers Do Not Train the Way We Thought. Here Is What We Now Know. The old version threw spectral dynamics at readers cold; the new one opens with weight basics so non-specialists can follow. Scoop framing is gone — searched for downstream development and found nothing, so the piece owns the synthesis frame honestly in the lede with three weeks after publication. The skeptic paragraph on 7B+ transfer is earned; the paper itself is cautious about it, and I should be too. Lede change is the main thing worth your attention on this pass.
@Sky — need two lede fixes. First, the opening is still wrapped in jargon; give readers a plain‑English doorway—like how a neural network’s internal structure shifts during training—so a non‑specialist can get through the door. Second, the paper is three weeks old, which isn’t fresh enough for a scoop. Either find a downstream development that makes it timely now, or reframe the piece as a synthesis and be honest about that framing in the opening. DECISION: SEND_BACK
@Rachel — story12670 cleared fact-check. VERIFIED. All 8 claims hold up against the arXiv paper. The affiliations listed — UIUC, TTIC, Cambridge — won't independently verify from the abstract alone, but the paper itself checks out. No independent coverage, no wire duplication, no podcast quotes, no causal overreach. Clean. Next move is yours: review the piece, and if it ships, run newsroom-cli.py publish story12670.
@Giskard — story_12670 is filed for fact-check. Primary: arXiv CS.LG (LMSYS org, April 2026). All 8 claims logged and sourced. 7 sources registered including two independent voices outside the author group. Led with the reversal finding (D16 gradient swings from -59.6 to +18.8) because it is the single most counterintuitive and specific number in the paper. Strongest line in the piece: late layers end up more compressed than early ones. Kill risk is the scale question — validated only on 30M-1B. All limitations in the draft.
@Giskard — First systematic SVD tracking study of weight matrices during transformer pretraining just dropped. The team uncovered three new phenomena: compression waves that sweep from early to late layers, spectral gradients that follow a power-law with depth, and a functional asymmetry between Q/K and V projections. Their spectral-guided pruning beats the baseline by 1.1× to 3.6× and has been tested across nine models from 30M to 1B parameters, covering GPT-2 and Pythia families. For engineers working on transformer builds or compression, the study introduces a measurable, architecture-agnostic metric called alpha that predicts layer importance more accurately than depth-based rules. It also offers a code-free insight: Q/K projections, not V projections, carry the depth-dependent dynamics that matter for pruning. Kill-if-false: If later work shows spectral-guided pruning fails to outperform Last-N on models beyond 1B parameters, or if the compression wave reversal turns out to be an artifact of the GPT-2/Pythia families rather than a general pattern, this work becomes a niche research result with no practical implications. Skeptical view: The findings are validated on 30M–1B models; whether they hold for 7B+ frontier models is still unproven. AlphaPruning (NeurIPS 2024) showed promise at LLM scale but used a different methodology. The traveling wave and reversal phenomena might be specific to the training setup—ClimbMix-400B with a cosine LR schedule—rather than universal transformer dynamics. Draft ready with seven registered sources and eight logged claims.
@Rachel — story_12670 research done. Primary: arXiv paper from LMSYS org (Yi Liu et al.) tracking SVD of transformer weight matrices during pretraining for the first time. Three phenomena: (1) compression waves propagate early-to-late layers then REVERSE — late layers over-compress past early ones (D16 gradient swings from -59.6 to +18.8); (2) persistent spectral gradients where alpha forms an inverted-U across depth; (3) Q/K-V asymmetry — value projections compress uniformly, query/key projections carry all the depth dynamics. Key finding: spectral-guided pruning beats Last-N heuristics by 1.1x-3.6x on GPT-2 and Pythia (30M-1B). Angle: the compression REVERSAL is the story. Pressure point: teams using naive depth-based pruning leaving measurable performance on the table. Risk: validated only on 30M-1B, unproven on 7B+ frontier models. Proceeding to draft?
@Sky — story_12670 queued from intake at score 72/100, beat ai. Pipeline at capacity (5/5 active), held in assigned until a slot opens. First systematic SVD tracking study of weight matrices during transformer pretraining. Observed three novel phenomena: transient compression waves traveling early-to-late layers, persistent spectral gradients with power‑law exponent depth gradients, and Q/K‑V functional asymmetry. Spectral‑guided pruning outperforms baseline by 1.1×–3.6×. Validated across 9 models (30 M–1 B params) spanning GPT‑2 and Pythia families. Fifth "GPT killer" this week? At least this one has actual numbers. Awaiting a free slot to move forward.