training-8b-multimodal-model-rl-find-scientific-errors-whole CLM-0648When training an 8B multimodal model with RL to find scientific errors in whole papers without being told where to look, mixing in matched easier variants of the same instances — where the error is asserted and the evidence supplied — yields better performance on the evidence-absent task itself than training only on that task at matched exposure.
Capability: Spotting an error in a given artifact
Observed on
Qwen3-VL-8B-Instruct, SFT then ~30 DAPO steps, on 12,900 paper-level error-detection samples built as matched Reason/Verify/Scan chains over six error categories; not tested on oth.
Sources
- Validated by: Compared against Pure Scan training (Scan samples only) at matched Scan exposure (step 10) and at longer training (step 20), and against a curriculum that switches to Scan-only after 15 steps; the staged setting scored higher on Scan composite reward (19.5 vs 16.8/16.9/16.5) and on completeness, alignment and precision components. All metrics come from the paper's own reward definitions, scored by Seed-1.6-Thinking as a fixed extractor.. Measured on one backbone with one RL algorithm and short training (10-30 steps); [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When a tool-using agent is trained with outcome-only RL and a surface prompt cue co-occurs with one tool in most training examples, the agent invokes that tool on tasks where it is useless, but only for tools it has already learned to use well and mainly when the cue is semantically related to the tool.Using the tools it is given · reviewed by AI
- In LLM fact-checking over supplied evidence, high veracity accuracy does not imply the verdict depends on that evidence: when the model's own cited sentences are iteratively removed, off-the-shelf models (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B) keep their SUPPORT/REFUTE labels at a non-zero plateau instead of abstaining, and training with paired supervision that requires NOT_ENOUGH_INFO on the evidence-ablated input removes most of that persistence without losing accuracy.Checking claims against evidence · reviewed by AI
- For 4-8B multimodal models on long-video QA, RL training on questions authored from short scene segments but answered over the whole video improves benchmark accuracy more than RL on questions authored from the whole video, because segment-authored questions target details a whole-video view overlooks; including the segment timestamp in the training instruction adds a further gain by removing referential ambiguity, and is not needed at inference.Reasoning about time in video · reviewed by AI
- When fine-tuning a 14B model on knowledge-graph-derived multiple-choice QA for a single medical domain, attaching supporting triples from the same source text chunk to each target triple during supervision, and repairing residual one-hop failures before reinforcement learning, gives higher accuracy on held-out 3-, 4- and 5-hop questions than training on isolated triples or running RL from an unrepaired checkpoint.Carrying a chain of inference through to its conclusion · reviewed by AI
- When a model is asked to audit many documents for planted errors in a single large-batch prompt, recall of known contaminants collapses and the model reports confident fabricated findings — invented absurd contaminants that appear in no document — rather than signalling incomplete processing.Stating false facts confidently · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: Training the same backbone only on evidence-absent samples, at equal or greater Scan exposure, matching or beating the staged mixture on evidence-absent error detection.