Influence functions unlock model steering through response rewriting
Researchers challenge a foundational assumption in training data attribution: that influence functions reliably identify examples worth modifying to steer model behavior. The work shows IF-selected samples often underperform random baselines under standard reweighting, suggesting the problem lies not with the examples themselves but with how they are intervened upon. By replacing responses with behavior-aligned or behavior-opposed supervision instead of reweighting, the team recovers the latent intervention value of influential samples across multiple open-weight LLMs. This reframes data attribution from a purely diagnostic tool into an actionable mechanism for model steering, with implications for both model debugging and adversarial robustness.
Modelwire context
ExplainerThe paper's core contribution isn't identifying influential samples (that's old), but showing that influence function rankings fail under reweighting specifically because reweighting is the wrong intervention lever. Replacing responses with explicit supervision recovers the signal, suggesting influence functions were always measuring something real about training dynamics, just not something that survives the standard application method.
This sits directly alongside the MemoryWalker paper from early September, which identified a training-inference mismatch in how agents compress context. Both papers share a pattern: the standard way practitioners apply a technique (reweighting for data attribution, linear context compression during training) creates a conditioning problem that obscures the underlying signal. The StateSwap work on hidden state framing effects also connects here, since both this paper and StateSwap suggest that how you intervene on a model matters as much as what you intervene on. The implication across all three is that mechanistic understanding of training dynamics is becoming prerequisite to making standard tools actually work.
If the same influential samples identified by influence functions show consistent intervention gains when responses are rewritten across different model families and scales (not just the open-weight LLMs tested here), that confirms this is a general property of how LLMs encode training signal rather than an artifact of specific architectures. If practitioners start adopting response-rewriting over reweighting for data steering in the next 6 months, that signals the field is moving from diagnosis to intervention.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsInfluence functions · Training data attribution · LLMs
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.